Portkey

No entries match these filters. Clear filters

1. changelog Oct 31, 2026 · issue 126

   Portkey adds a pre-model traffic inspection lifecycle stage and expands guardrail coverage to MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. tool calls⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

   • Portkey now applies guardrails to MCP tool calls, extending coverage beyond standard model request/response traffic.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Portkey adds a new lifecycle stage to inspect traffic before it reaches a model.

   2 details in Everything

2. changelog Oct 16, 2026 · issue 110

   Portkey adds Decisions API, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. guardrails, startHooks lifecycle stage, and Meta/ElevenLabs provider support⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

   └──▷ TRY IT
   Call the Decisions API to get a calibrated boolean judgment from a TypeSafe Jev model — useful for automated triage or policy enforcement pipelines.
   copy

   $ curl -X POST https://api.portkey.ai/v1/decisions \
     -H 'x-portkey-api-key: <your-api-key>' \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "jev",
       "question": "Does this request contain PII?",
       "state": {"text": "My name is Jane Doe and my SSN is 123-45-6789"},
       "output_type": "boolean"
     }'

   • Adds startHooks lifecycle stage to screen and normalize traffic before authentication, routing, or other hooks run — including two new checks: Header Check (allow/block by header key-value pairs) and Header Transform (remove, set, or rewrite headers before routing; x-portkey-* headers are protected); existing checks allowedRequestTypes, webhook, and JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. header lookup can also now run as startHooks.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Adds a Decisions API endpoint for typed-judgment models: send a question with structured state and receive a typed judgment (boolean, choice, or score) with probabilities and a calibrated confidence score, routing through the full gateway stack (auth, routing, retries, caching, logging, cost tracking); initial support covers TypeSafe's Jev model family.
   • Extends MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. guardrails so MCP tool calls pass through the same pre- and post-execution checks available for LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. requests; denied calls return a structured error with the guardrail and check ID.
   • Adds per-attempt retry logging, giving each retry its own log row with its own status and response instead of only surfacing the final result.
   • Adds Meta and ElevenLabs as new providers: Meta supports Chat Completions, Messages API, Responses API, and Files API, while ElevenLabs supports text-to-speech and speech-to-text with full observability and reliability features.

   +5 more show less

   • Extends xAI provider to support image generation with OpenAI-compatible parameters plus xAI-specific aspect ratio and resolution controls.
   • Adds Anthropic-protocol auto-detection on /v1/models so Claude Code and other Anthropic-native clients receive Anthropic's native model shape instead of the OpenAI-normalized shape.
   • Extends JWT authentication to accept multiple JWKSJWKSA JSON Web Key Set, a published collection of cryptographic keys. Applications use it to find the public key needed to verify a token's signature. URLs and attribute requests to the matching workspace user.
   • Extends conditional routing to read fields from multipart/form-data requests (image edits, audio transcriptions).
   • Messages API reasoning models now stream their thinking phase as a thinking content block.
   • Adds startHooks lifecycle stage to screen and normalize traffic before authentication, routing, or other hooks run — including two new checks: Header Check (allow/block by header key-value pairs) and Header Transform (remove, set, or rewrite headers before routing; x-portkey-* headers are protected); existing checks allowedRequestTypes, webhook, and JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. header lookup can also now run as startHooks.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Adds a Decisions API endpoint for typed-judgment models: send a question with structured state and receive a typed judgment (boolean, choice, or score) with probabilities and a calibrated confidence score, routing through the full gateway stack (auth, routing, retries, caching, logging, cost tracking); initial support covers TypeSafe's Jev model family.
   • Adds per-attempt retry logging, giving each retry its own log row with its own status and response instead of only surfacing the final result.
   • Adds Meta and ElevenLabs as new providers: Meta supports Chat Completions, Messages API, Responses API, and Files API, while ElevenLabs supports text-to-speech and speech-to-text with full observability and reliability features.

   6 more in Everything

3. docs update Oct 8, 2026 · issue 097

   Portkey adds Mistral AI content moderation guardrails with 9 policy categories and multilingual support across 11 languages.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

   └──▷ USE IT
   Apply Mistral content moderation to both incoming prompts and outgoing LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. responses by wiring guardrail IDs into a Portkey Config.
   json copy

   {
     "input_guardrails": ["guardrails-id-xxx"],
     "output_guardrails": ["guardrails-id-xxx"]
   }

   Send a chat completion through Portkey with a Mistral moderation Config attached, so every request and response is checked before delivery.
   copy

   $ curl https://api.portkey.ai/v1/chat/completions \
     -H 'Content-Type: application/json' \
     -H 'Authorization: Bearer $OPENAI_API_KEY' \
     -H 'x-portkey-api-key: $PORTKEY_API_KEY' \
     -H 'x-portkey-config: $CONFIG_ID' \
     -d '{"model": "gpt-3.5-turbo", "messages": [{"role": "user", "content": "Hello!"}]}'

   Connect your Mistral La Plateforme API key to Portkey and create a Moderate Content guardrail covering PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. and Violence categories before attaching it to a Config.
   📍1. In Portkey, go to Settings › Integrations, click Edit on the Mistral integration, and paste your Mistral La Plateforme API Key. 2. Go to Guardrails › Create, search for Moderate Content, click Add, select the desired Moderation Checks (e.g. PII, Violence and Threats), set Timeout to 5000ms, and save to obtain a Guardrail ID. 3. Create a Config with that Guardrail ID in input_guardrails or output_guardrails and save it to obtain a Config ID for use in requests.
   • Adds input_guardrails and output_guardrails config params to route requests through Mistral's Moderate Content guardrail check, supporting beforeRequestHook and afterRequestHook.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Adds Mistral Moderate Content guardrail check with configurable Moderation Checks (array) and Timeout (number, default 5000ms) parameters to detect harmful content across 9 policy categories: Sexual, Hate and Discrimination, Violence and Threats, Dangerous and Criminal Content, Self-harm, Health, Financial, Law, and PII.
   • Supports Mistral's natively multilingual moderation across Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

   3 details in Everything

4. docs update Oct 8, 2026 · issue 097

   Portkey adds Lakera Guard integration for prompt/response screening with PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. redaction via /v2/guard⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

   └──▷ USE IT
   Attach a Lakera Guard guardrail to both input and output of a Portkey config to screen all prompts and responses for policy violations and PII.
   json copy

   {
     "input_guardrails": ["guardrails-id-xxx"],
     "output_guardrails": ["guardrails-id-yyy"]
   }

   Set up the Lakera integration credentials and configure the Guard check with a specific project and timeout via Admin Settings.
   📍1. Go to Admin Settings → Plugins 2. Edit the Lakera integration, paste your Lakera API Key (from platform.lakera.ai), and optionally set API Base URL to https://api.lakera.ai 3. Go to Guardrails → Create, search for Lakera Guard, click Add 4. Set projectID to your Lakera project, endInclusive to false, and timeout as needed 5. Set guardrail actions and click Create
   • Adds Lakera Guard integration that screens prompts and responses via Lakera's /v2/guard API, blocking on policy hits and redacting PII spans when pii/* detectors fire.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Exposes projectID, endInclusive, and timeout parameters on the Lakera Guard check, supporting both beforeRequestHook and afterRequestHook positions.
   • Attaches Lakera guardrails to Portkey configs via input_guardrails and output_guardrails fields for SDK and gateway use.
   • Allows an optional API Base URL override on the Lakera integration (defaults to https://api.lakera.ai) configured under Admin Settings → Plugins.

   4 details in Everything

5. docs update Oct 6, 2026 · issue 096

   Portkey adds a /v1/decisions endpoint for typed model judgments (choice, score, probability) with guardrails on the state field⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

   └──▷ TRY IT
   Route an incoming support ticket to the right team by getting a structured department classification instead of parsing free text.
   copy

   $ curl https://api.portkey.ai/v1/decisions \
     -H 'x-portkey-provider: typesafe' \
     -H 'Authorization: Bearer $TYPESAFE_API_KEY' \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "jev-latest",
       "state": "My payouts have been failing for 3 days.",
       "questions": {
         "department": {
           "type": "choice",
           "instructions": "Which team should handle this?",
           "criteria": { "billing": "Payments", "technical": "Bugs" }
         }
       }
     }'

   Prioritise a user's own workspace when issuing JWTsJWTsJSON Web Tokens — an open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object. Cyber tools use them for stateless authentication without server-side session storage., so multi-tenant deployments resolve to the correct workspace without relying on deployment-level defaults.
   copy

   $ JWT_PREFER_USER_WORKSPACE_RESOLUTION=ON ./portkey-gateway

   • Adds POST /v1/decisions endpoint that returns typed judgments (noul, choice, or score) instead of free text, accepting a state field and a questions map — currently available through TypeSafe's Jev models (jev-latest, jev-1.13.0).⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
   • Adds JWT_PREFER_USER_WORKSPACE_RESOLUTION environment variable to the gateway; set to ON to try a user token's own workspace membership before deployment and organisation defaults in the resolution order.
   • Adds guardrails support for /v1/decisions requests, scanning only the caller-supplied state field (never model, questions, or criteria) via input_guardrails at the before-request stage — hard-deny returns 446, soft-deny returns 200 with an empty answers map.
   • Decisions requests participate in the full gateway feature set: retries, fallbacks, load balancing, and circuit breaking via Configs; simple and semantic caching; and observability (logs, cost tracking, analytics).
   • Mutator guardrails (e.g. PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. redaction) can rewrite state before it reaches the provider, with JSON round-trip safety — if redacted output is no longer valid JSON for an object/array state, Portkey preserves the original rather than forwarding corrupted content.

   5 details in Everything

6. Enterprise Gateway 2.26.0 Oct 1, 2026 · release notes

   DashScope image generation support, user-first workspace resolution via JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store., and Gemini finish_reason reporting.

   • Adds DashScope as a supported provider on the /v1/images/generations endpoint with region-based pricing attribution in logs.
   • Adds JWT_PREFER_USER_WORKSPACE_RESOLUTION=ON environment variable to resolve workspace defaults from user membership when none is named in the JWT.
   • Improves Google Gemini / Vertex AI streaming responses to report finish_reason: "tool_calls" for tool invocations, matching non-streaming behavior.

   3 details in Everything

7. Enterprise Gateway 2.25.1 Sep 29, 2026 · release notes

   Portkey Enterprise Gateway 2.25.1 adds Meta provider support, new Responses API operations, and improved MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. auth handling.

   • Adds Meta as a first-class provider with support for Chat Completions, Anthropic-compatible Messages API, OpenAI-compatible Responses API, and Files API protocols.
   • Introduces cancel and input token counting operations in the Responses API for supported providers.
   • Wires Retrieve, delete, and list input items endpoints for xAI, DashScope, Perplexity, and Azure OpenAI providers via the Responses API.
   • Unstrips interleaved-thinking-2025-05-14 and per-turn-control-2026-07-01 beta headers for Anthropic on Bedrock and Vertex AI.
   • Streams thinking phase as a thinking content block in the Messages API instead of dropping it for routed reasoning models.

   +1 more show less

   • Enables 401 challenge and immediate re-authorization for revoked tokens in the MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. during mid-refresh.

   6 details in Everything

8. Backend 1.28.0 Sep 28, 2026 · release notes

   Portkey Backend v1.28.0 improves log export filtering with deployment_slug and optimizes list exports response payload. Model catalog definitions and pricing are updated for Anthropic, Bedrock, Vertex AI, Google, Azure AI, OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code., Lightning AI.

9. docs update Sep 24, 2026 · issue 084
   └──▷ TRY IT
   Generate speech via ElevenLabs through Portkey's unified audio endpoint without any custom host setup.
   copy

   $ curl -X POST https://api.portkey.ai/v1/audio/speech \
     -H "x-portkey-provider: @YOUR_ELEVENLABS_PROVIDER_SLUG" \
     -H 'Content-Type: application/json' \
     -d '{"model": "eleven_multilingual_v2", "input": "Hello! Welcome to Portkey.", "voice": "21m00Tcm4TlvDq8ikWAM"}'

   Transcribe audio using ElevenLabs Scribe via Portkey's unified transcriptions endpoint.
   copy

   $ curl -X POST https://api.portkey.ai/v1/audio/transcriptions \
     -H "x-portkey-provider: @YOUR_ELEVENLABS_PROVIDER_SLUG" \
     -F file="@recording.mp3" \
     -F model="scribe_v1"

10. docs update Sep 24, 2026 · issue 084

    Portkey adds native ElevenLabs TTS and STT routing via unified /v1/audio/speech and /v1/audio/transcriptions endpoints⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

    └──▷ TRY IT
    Generate speech via ElevenLabs through Portkey's unified audio endpoint without any custom host setup.
    copy

    $ curl -X POST https://api.portkey.ai/v1/audio/speech \
      -H "x-portkey-provider: @YOUR_ELEVENLABS_PROVIDER_SLUG" \
      -H "Content-Type: application/json" \
      -d '{"model": "eleven_multilingual_v2", "input": "Hello! Welcome to Portkey.", "voice": "21m00Tcm4TlvDq8ikWAM"}'

    Transcribe audio using ElevenLabs Scribe through Portkey's unified transcription endpoint.
    copy

    $ curl -X POST https://api.portkey.ai/v1/audio/transcriptions \
      -H "x-portkey-provider: @YOUR_ELEVENLABS_PROVIDER_SLUG" \
      -F file="@recording.mp3" \
      -F model="scribe_v1"

    • Routes ElevenLabs text-to-speech through the unified /v1/audio/speech endpoint using x-portkey-provider: @YOUR_ELEVENLABS_PROVIDER_SLUG, accepting any ElevenLabs voice ID (premade, cloned, or custom) via the voice field — forwarded as-is rather than remapped from OpenAI built-in voice names.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
    • Routes ElevenLabs speech-to-text through the unified /v1/audio/transcriptions endpoint using the scribe_v1 model, with no custom host configuration required.
    • Adds providerSlug as an opt-in Prometheus extra label (disabled by default; enable by adding providerSlug to PROMETHEUS_EXTRA_LABELS) to distinguish the Model Catalog provider or virtual key that served a request from the upstream provider name.

    3 details in Everything

11. Enterprise Gateway 2.25.0 Sep 24, 2026 · release notes

    Enterprise Gateway 2.25.0 adds TypeSafe Decisions API, workspace selection via JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. header, and startHooks for guardrails

    • Adds /v1/decisions endpoint for typed-judgment models returning boolean, choice, or score with probabilities
    • Adds x-portkey-workspace header support to gateway-local JWT requests to select a workspace at request time
    • Adds startHooks list in guardrail configs to run checks before authentication and routing
    • Adds Header Check and Header Transform guardrail checks available at the startHooks stage

    4 details in Everything

12. Enterprise Gateway 2.24.0 Sep 22, 2026 · release notes

    Enterprise Gateway 2.24.0 adds per-attempt retry logging, truncated SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. detection across providers, and Anthropic safeguard forwarding.

    • Adds retry_count, response_status, response_headers, and response_body fields to log rows for intermediate retry attempts
    • Enables stream_incomplete error event appending when a stream ends without its terminal event across all unified text routes (/chat/completions, /completions, /messages, /responses)
    • Forwards the safeguards and safeguard_results fields to Anthropic on Claude Code auto-mode requests

    3 details in Everything

13. Backend 1.27.0 Sep 21, 2026 · release notes

    Backend 1.27.0 adds SCX.ai provider, per-model pricing multipliers, guardrails analytics endpoints, Azure PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. category filtering, and new API key filters.

    • api — Adds /guardrail/summary, /guardrail/trend, and /guardrail/grouped endpoints for Guardrails analytics.
    • config — Adds piiCategories parameter to Guardrails settings to restrict detection to selected PII categories.
    • api — Adds filter parameters to the API key list endpoint.
    • config — Allows setting pricing multipliers on individual models within an integration.
    • New scx-ai provider support (OpenAI-compatible) in the Model Catalog.

    5 details in Everything

14. Enterprise Gateway 2.23.0 Sep 18, 2026 · release notes

    Portkey Enterprise Gateway 2.23.0 adds ElevenLabs support, Vertex AI region honor, Prometheus provider label, and OAuth error exposure

    • Adds native routing for /v1/audio/speech and /v1/audio/transcriptions to ElevenLabs without custom host configuration
    • Makes vertexRegion config key honored for Vertex AI Rerank requests instead of defaulting to global Discovery Engine
    • Adds opt-in Prometheus metrics label providerSlug to identify the Model Catalog provider that served a request
    • Exposes underlying OAuth authorize error on rejection via MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. instead of returning HTTP 500

    4 details in Everything

15. Frontend 1.10.1 Sep 17, 2026 · release notes

    Frontend 1.10.1 adds guardrail workspace exclusions, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface./A2AA2AAgent2Agent, an open protocol started by Google for AI agents from different vendors to discover each other, delegate tasks and exchange results. It complements MCP, which connects an agent to tools. grouping and deletion, and API key pagination

    • Adds UI for organisation-level guardrail workspace exclusions in input and output guardrails
    • Adds MCP and A2A grouping in the integration management UI
    • Adds delete option for MCP servers in the dashboard
    • Adds pagination for API keys listing

    4 details in Everything

16. docs update Sep 16, 2026 · issue 076

    Portkey MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. gains two-layer authentication with API key, OAuth 2.1OAuth 2.1A consolidated draft of the OAuth authorization framework that merges OAuth 2.0 with its most important security BCP updates, dropping unsafe flows so tools can delegate user access without handling credentials directly., external IdPIdPA service that authenticates users and supplies identity information to other applications. It is the sign-in authority behind many single sign-on systems., and per-model pricing adjustments⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

    • Adds MCP Gateway authentication via x-portkey-api-key header (API key scoped to workspaces, requires mcp invoke permission, created in Settings > API Keys).⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠
    • Adds OAuth 2OAuth 2An open authorization framework (RFC 6749) that lets a service grant limited access to a user's resources on another service without exposing credentials, giving cyber tools a standard way to authenticate against third-party APIs..1 support for MCP Gateway: when no API key is present Portkey acts as the OAuth provider and initiates a browser-based consent flow for interactive tools such as Claude Desktop or Cursor.
    • Adds External IdP (Bring Your Own Identity) support for MCP Gateway, enabling JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. validation from Okta, Auth0, Azure AD, and Cognito so users without Portkey accounts can access MCP servers.
    • Adds three MCP server authentication methods when configuring an upstream server: OAuth Auto (full OAuth 2.1 with dynamic client registration and automatic token refresh), Client Credentials (shared OAuth client ID/secret with auto-refresh), and Custom Headers (static API keys or tokens via configurable headers).
    • Adds identity forwarding from Portkey to upstream MCP servers after gateway authentication, supporting Claims Header (JSON object of user claims), Bearer Passthrough (original token), and Signed JWT (Portkey-signed, cryptographically verifiable JWT).

    +3 more show less

    • Adds header forwarding from agent requests to MCP servers for distributed tracing, tenant context, or custom metadata.
    • Adds per-model pricing_adjustments on individual models within an Integration (requires Gateway version 2.22.0 or higher); a model-level adjustment fully replaces the Integration-level one for that model.
    • Adds CIE Directory Sync for SCMSCMSource Code Management: version-control systems (Git, SVN, etc.) that track changes to code over time, giving cyber tools reproducible builds, audit trails, and collaborative development workflows. (Strata Cloud Manager) deployments, enabling automatic workspace provisioning by mapping Palo Alto Networks Cloud Identity Engine groups — sourced from Entra IDEntra IDMicrosoft's cloud-based identity and access management service, formerly Azure Active Directory, used by cyber tools to authenticate users and enforce access policies in Microsoft 365 and Azure environments., Okta, or on-premises Active DirectoryActive DirectoryMicrosoft's directory service for Windows networks that centralises authentication and authorisation; cyber tools target or integrate with it because it controls identity and access across most enterprise environments. — to AI Gateway workspaces.

    8 details in Everything

17. docs update Sep 15, 2026 · issue 075

    Portkey's Azure Guardrails integration adds PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. category filtering to restrict detection to specific entity types.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

    ⤢
    • Adds PII category filtering to Azure Guardrails, supporting entity types Person, Email, PhoneNumber, Address, CreditCardNumber, and PassportNumber — leave empty to detect all default categories.⁠︀︄︇︀︅︉︀︅︀︀︃︀︀︅︅︀︅︊︀︄︆︀︃︄⁠

    1 detail in Everything

18. Frontend 1.10.0 Sep 11, 2026 · release notes

    Portkey Frontend 1.10.0 adds Meta provider, OAuth UI for OpenAI, Guardrail analytics dashboard, and custom legacy model pricing.

    • Adds OAuth Client Credentials UI for OpenAI integrations
    • Adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Guardrails support
    • Adds custom pricing for legacy models support

    3 details in Everything

19. Enterprise Gateway 2.22.0 Sep 11, 2026 · release notes

    Azure context-aware pricing, per-model pricing adjustments, and Claude Code model discovery via /v1/models

    • Adds id, type, display_name, and created_at fields in the response body for /v1/models calls from clients sending an anthropic-version header.
    • Adds support for the service_tier request parameter to forward Azure service tiers (standard, batch, priority, flex) to providers, replacing dropped values.
    • Adds PII Categories setting in Azure PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. Guardrail configurations to limit detection to specific entity types.

    3 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Automatic anthropic-beta: messages-2023-12-15 header is no longer added to Anthropic requests.
    • !A model-level pricing adjustment fully replaces the integration-level one rather than merging with it.
20. Backend 1.26.0 Sep 6, 2026 · release notes

    Portkey Backend 1.26.0 adds ElevenLabs provider support, multi-JWSS JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. validation, webhook proxy guardrails, and log export API.

    • Adds elevenlabs provider to the Model Catalog for text-to-speech, speech-to-text, dubbing, and voice cloning
    • Allows JWT validation to accept comma-separated JWKSJWKSA JSON Web Key Set, a published collection of cryptographic keys. Applications use it to find the public key needed to verify a token's signature. URIs for multiple identity providersidentity providersA service that authenticates users and supplies identity information to other applications. It is the sign-in authority behind many single sign-on systems., fetching keys in parallel
    • Adds executeOnProxy parameter to webhook guardrails to run on /v1/* pass-through requests with method, path, and content-type metadata
    • Adds DELETE /v1/logs/exports/{id} endpoint to delete log exports (requires cancellation of in-progress exports first)
    • Adds default timeout field for AIRS service requests

    +2 more show less

    • Routes model fetching for SCMSCMSource Code Management: version-control systems (Git, SVN, etc.) that track changes to code over time, giving cyber tools reproducible builds, audit trails, and collaborative development workflows. through the config proxy
    • Returns usage policy IDs in organisation data responses via the Organisation sync API

    7 details in Everything

21. Enterprise Gateway 2.21.0 Sep 4, 2026 · release notes

    ElevenLabs provider support, Gateway-Local JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. multi-JWKSJWKSA JSON Web Key Set, a published collection of cryptographic keys. Applications use it to find the public key needed to verify a token's signature./attribute, and SCX.ai provider added in Portkey Enterprise Gateway v2.21.0

    • Adds ElevenLabs as a new provider for text-to-speech and speech-to-text requests routed through the gateway
    • Gateway-local JWT auth now accepts a comma-separated list of JWKS URLs via jwks_urls for multi-identity validation
    • Requests authenticated via Gateway-local JWT are attributed to existing Portkey users when their email matches a workspace user
    • Adds SCX.ai as a new OpenAI-compatible provider for chat completions

    4 details in Everything

22. Frontend 1.9.6 Sep 2, 2026 · release notes

    Portkey Frontend 1.9.6 adds Deepgram and Singulr integrations, OCROCROptical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels. model support, Owner role, configurable API key scopes, and analytics chart APIs.

    • Adds Deepgram as a new provider in the dashboard
    • Adds OCR model type support in the Model Catalog UI
    • Adds Singulr guardrail partner logo and plugin UI
    • Adds Headroom plugin
    • Adds Owner role support for administration

    +1 more show less

    • Adds tokens and request chart token breakup API integration

    6 details in Everything

23. Enterprise Gateway 2.20.0 Sep 1, 2026 · release notes

    Enterprise Gateway 2.20.0 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. guardrails, unified gateway/MCP mode via SERVER_MODE=unified, and proxy webhook guardrails.

    • MCP tool calls can now be checked against guardrails before and after execution.
    • Set SERVER_MODE=unified to run the AI Gateway and MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. on a single port.
    • Webhook guardrail can now run on proxy (/v1/*) requests via an opt-in executeOnProxy parameter.

    3 details in Everything

24. Backend 1.25.0 Sep 1, 2026 · release notes

    Backend 1.25.0 introduces MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. guardrails, WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. auth, deployment tags, and expanded Model Catalog integrations.

    • Adds capability_ids mappings and run_on values (beforeToolCall, afterToolCall) to attach guardrails to MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. servers
    • Adds new MCP registry endpoint for listing available MCP servers with cursor-based pagination
    • Integrates new partner guardrails for AI security scanning (singulr) and Alibaba Cloud's content moderation service
    • Adds Deepgram (deepgram), Lightning AI (lightning-ai), and Headroom (headroom) to the Model Catalog
    • Adds workload identity federation (wif) as a provider authentication type for keyless cloud workload authentication

    +9 more show less

    • Adds tags (key-value pairs) support for filtering and organisation in Deployments
    • Adds source_type field and MCP Gateway base URL configuration support in Deployments
    • Removes default target filter from the Guardrail list API, returning both LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. and MCP guardrails
    • Adds hook check data (guardrail check results and errors) to generation log responses
    • Adds validation of guardrail slugs at save time for input_guardrails and output_guardrails
    • Adds support for the forwardMetadata parameter in F5 Guardrails
    • Adds log export deletion support via the API
    • Harden API key scope enforcement so users cannot create keys with scopes exceeding their role permissions
    • Adds service-role support for Bedrock plugins in self-hosting configurations

    14 details in Everything

25. docs update Aug 31, 2026 · issue 060

    Portkey MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. adds Gmail integration for reading, sending, and managing email from AI agents via OAuth 2.1OAuth 2.1A consolidated draft of the OAuth authorization framework that merges OAuth 2.0 with its most important security BCP updates, dropping unsafe flows so tools can delegate user access without handling credentials directly.

    └──▷ USE IT
    Wire a Gmail MCP server into an agent so it can search and send email through Portkey's managed gateway.
    json copy

    {
      "gmail": {
        "url": "https://mcp.portkey.ai/gmail/mcp"
      }
    }

    Register Gmail in Portkey Advanced Configuration with the OAuth credentials from your Google Cloud app, scoped to read and send only.
    json copy

    {
      "client_id": "YOUR_GOOGLE_CLIENT_ID",
      "client_secret": "YOUR_GOOGLE_CLIENT_SECRET",
      "scope": "https://www.googleapis.com/auth/gmail.readonly https://www.googleapis.com/auth/gmail.send"
    }

    • Adds https://mcp.portkey.ai/gmail/mcp as a hosted MCP Gateway endpoint for Gmail, enabling AI agents to read, send, and manage email without running a self-hosted binary.
    • Supports Gmail OAuth 2OAuth 2An open authorization framework (RFC 6749) that lets a service grant limited access to a user's resources on another service without exposing credentials, giving cyber tools a standard way to authenticate against third-party APIs..1 with manual OAuth metadata — configure client_id, client_secret, redirect_uri, and scope in Portkey Advanced Configuration using slug gmail and server URL https://gmailmcp.googleapis.com/mcp/v1.
    • Exposes Gmail MCP tools list_messages (with maxResults, labelIds, q, pageToken parameters) and get_message (with id and format parameters) for agent-driven mailbox operations.
    • Supports fine-grained Gmail OAuth scopes: gmail.readonly, gmail.send, gmail.modify, gmail.compose, and gmail.labels — agents only receive access to configured scopes.
    • Portkey stores and automatically refreshes Gmail OAuth tokens after first-use authorization, so agents do not need to re-authenticate on subsequent calls.

    5 details in Everything

26. docs update Aug 22, 2026 · issue 053

    Portkey adds a Registry API for approved MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. servers, configurable semantic cache embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. providers, and OAuth metadata controls for MCP integrations.

    └──▷ TRY IT
    Fetch your organization's full approved MCP server catalog for use with a governance-aware MCP client that restricts users to sanctioned servers.
    copy

    $ curl https://mcp.portkey.ai/v0.1/servers \
      -H "x-portkey-api-key: $PORTKEY_API_KEY"

    Register a confidential Google OAuth App in Portkey MCP to obtain refresh tokensrefresh tokensA credential used to obtain a new access token without signing in again. It is sent to the authorization service rather than used directly to call the protected API. so sessions survive access tokenaccess tokenA credential that authorizes a client to use a service or API. It usually has limited permissions and a lifetime; anyone holding it may be able to use those permissions. expiry.
    json copy

    {
      "client_id": "...apps.googleusercontent.com",
      "client_secret": "your_google_oauth_client_secret",
      "scope": "https://www.googleapis.com/auth/drive.readonly",
      "authorization_params": {
        "access_type": "offline",
        "prompt": "consent"
      }
    }

    • Adds GET https://mcp.portkey.ai/v0.1/servers Registry API (beta) serving approved MCP servers as a standards-compliant server.json catalog in MCP Registry v0.1 format, authenticated via x-portkey-api-key header with mcp_servers.list scope.
    • Adds query parameters limit (1–100), cursor, search, updated_after (ISOISOA widely adopted international standards body that publishes technical specifications; cyber tools cite ISO standards to signal compliance with recognized benchmarks for quality, security, or interoperability. 8601), and workspace_id (UUIDUUIDA universally unique identifier, usually a 128-bit value used to identify an object across systems. It is designed to make independently generated identifiers unlikely to collide. or slug) to the GET /v0.1/servers endpoint for pagination, filtering, and incremental syncs.
    • Adds SEMANTIC_CACHE_EMBEDDING_PROVIDER environment variable supporting openai, azure-openai, google (Gemini embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.), and vertex-ai to configure which provider generates embeddings for semantic caching.
    • Adds SEMANTIC_CACHE_EMBEDDINGS_URL, SEMANTIC_CACHE_EMBEDDING_MODEL, SEMANTIC_CACHE_EMBEDDING_DIMENSIONS, and SEMANTIC_CACHE_EMBEDDING_API_KEY environment variables to configure the chosen embedding provider for semantic cache.
    • Adds client_id, client_secret, and authorization_params fields to the OAuth Metadata configuration in the MCP Registry, allowing pre-registered OAuth clients and disabling Dynamic Client Registration.

    +3 more show less

    • Adds token_endpoint_auth_method field to OAuth metadata, accepting client_secret_post or client_secret_basic for confidential OAuth app clients (default is none).
    • Adds authorization_params object to OAuth metadata for appending extra string-valued query parameters to upstream authorization URLs, enabling provider-specific flows such as Google's access_type=offline and prompt=consent for refresh tokenrefresh tokenA credential used to obtain a new access token without signing in again. It is sent to the authorization service rather than used directly to call the protected API. acquisition.
    • Exposes Portkey's OAuth callback URL at https://mcp.portkey.ai/oauth/upstream-callback (or <MCP_GATEWAY_BASE_URL>/oauth/upstream-callback for self-hosted deployments) for MCP OAuthMCP OAuthAn authentication layer for the Model Context Protocol that lets MCP clients obtain delegated access tokens so cyber tools can call MCP servers on a user's behalf without handling raw credentials. App registration.

    8 details in Everything

27. docs update Aug 22, 2026 · issue 053

    Portkey Prometheus metrics gain PROMETHEUS_EXTRA_LABELS to expose API key and config name as metric labels.

    └──▷ USE IT
    Break down Prometheus metrics by Portkey API key name and gateway config name to isolate costs or latency per team or workflow.
    .env copy

    PROMETHEUS_EXTRA_LABELS=apiKeyName,configName

    • Adds PROMETHEUS_EXTRA_LABELS environment variable (comma-separated string, default empty) to attach apiKeyName and/or configName as labels on every custom Prometheus metric, enabling per-key and per-config filtering in dashboards.

    1 detail in Everything

28. docs update Aug 22, 2026 · issue 053

    Portkey adds MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once., Agent Gateway, Coding Agent governance, gRPCgRPCAn open-source remote procedure call framework from Google that uses HTTP/2 and Protocol Buffers to let services call each other's functions across a network, giving tools low-latency, strongly typed inter-process communication. transport, OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. observability, and Nitro Mode across its AI gateway platform.

    • New MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Gateway provides centralized authentication, access control, and governance for MCP servers, with a Registry API (Beta) to serve approved servers as a standard endpoint.
    • Adds OAuth 2.1OAuth 2.1A consolidated draft of the OAuth authorization framework that merges OAuth 2.0 with its most important security BCP updates, dropping unsafe flows so tools can delegate user access without handling credentials directly. built-in authentication (OAuth), external identity-provider support (Bring Your Own Auth), JWT Validation, Identity Forwarding, Forwarding Headers, Custom Auth, and Authorization controls for MCP Gateway.
    • Adds Tool Provisioning to control which tools, resources, and prompts are exposed to which workspaces and users through the MCP Gateway.
    • Adds Team Provisioning (access control) to control which workspaces and users can access specific MCP servers.
    • Adds Rate Limits and Circuit Breakers for MCP tool call requests and token consumption within the MCP Gateway.

    +30 more show less

    • Supports adding internal and external MCP servers (e.g. Linear) to a managed MCP Registry with add/manage/govern workflows.
    • New Remote MCP support in the AI Gateway routes requests to MCP servers directly through the gateway.
    • New Agent Gateway provides centralized authentication, access control, and observability for AI agents.
    • New Coding Agents feature adds a governance layer for Claude Code, Codex, and similar coding agents.
    • New gRPC (Beta) transport lets the AI Gateway accept gRPC as an alternative protocol to HTTP.
    • New Nitro Mode (Beta) forwards the request body directly to the provider, bypassing gateway transformation for maximum performance.
    • Adds OpenTelemetry integration for LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. observability, supporting any OpenTelemetry-compatible instrumentation library.
    • New Auto-Instrumentation [BETA] enables automatic tracing of LLM calls without manual SDK instrumentation.
    • New Open Responses endpoint supports the Responses API format with any LLM provider.
    • New Messages endpoint supports Anthropic's Messages API format with any provider through the Universal API.
    • New Unified Batch Inference (batches) runs large-scale inference jobs through one API across providers, with Guardrails support via Guardrails for Batches.
    • New Files API allows uploading files to Portkey and reusing content across requests.
    • New Fine-tuning support runs fine-tuning jobs through the Portkey Gateway.
    • New Circuit Breaker in the AI Gateway automatically stops routing to unhealthy targets.
    • New Model Catalog provides a single pane to view and manage every AI provider integration, including Integrations (credential store), Adding Custom Models, Adding Pricing Adjustments, and Overriding Model Details.
    • Adds Portkey Models as an open-source pricing and configuration database for LLM models.
    • Adds Budget Limits to set cost limits per integration/provider key.
    • Adds Guardrails for Embedding Requests to apply security and data validation to embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. API calls.
    • Adds PII Redaction guardrail to replace sensitive data in requests before they reach providers.
    • Adds Creating Raw Guardrails (in JSON) mode to define guardrails directly as JSON config.
    • Adds Forwarding Headers for guardrails to pass client request headers through to guardrail services.
    • Adds Bring Your Own Guardrails integration support for custom guardrail providers.
    • New Prompt Partials let practitioners save reusable prompt components and compose them across templates.
    • Adds Sharing Prompts and Partials to share prompt templates and partials across workspaces.
    • Adds Logs Export to access Portkey log data for external analysis.
    • Adds Feedback APIs to attach human or automated feedback signals to logged LLM requests.
    • Adds Metadata support to attach custom context key-value pairs to AI requests for filtering and analysis.
    • Adds Connect Bedrock with Amazon Assumed Role integration for IAMIAMIdentity and Access Management — a framework of policies and services for controlling who can authenticate to systems and what resources they are authorized to use, letting cyber tools enforce least-privilege and audit access. role-based Bedrock authentication.
    • Adds Canary Testing to split traffic across LLM variants for safe rollouts.
    • Adds Realtime API support to use OpenAI's Realtime API with logs and cost tracking through the gateway.

    35 details in Everything

29. Enterprise Gateway 2.19.0 Aug 21, 2026 · release notes

    Portkey Enterprise Gateway 2.19 adds Deepgram, OAuth client credentials, Vertex AI OCROCROptical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels., MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Registry Proxy and cost tracking for Meshy/Tripo3D.

    • Adds Deepgram provider integration for speech-to-text and text-to-speech with full observability
    • Allows OpenAI-compatible integrations to authenticate to upstream providers using OAuth Client Credentials grant (RFC 6749) instead of static API keys
    • Extends /v1/ocr endpoint to support Mistral OCR models on Google Vertex AI with per-page pricing
    • Adds Responses and Messages endpoints support for the Qwen provider, along with additional non-OpenAI parameters
    • Introduces Headroom plugin to compress request context before model ingestion, configured via standard guardrails workflow

    +4 more show less

    • Uses provider-reported cost for billing attribution when available (starting with OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code.)
    • Enables MCP Registry Proxy in the MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. to expose registry endpoints like server listings directly through the gateway
    • Adds credit-based cost tracking for Meshy and Tripo3D 3D-generation providers with a 24h duplicate-billing protection window
    • Supports per-request AWS Bedrock STSSTSAWS Security Token Service, an Amazon Web Services API that issues short-lived credentials for IAM roles, letting cyber tools authenticate to AWS without storing long-term secrets. session tags via the AWS_BEDROCK_STS_SESSION_TAGS_ENABLED environment variable

    9 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Removed the deprecated temperature handling for Claude 4 models in Amazon Bedrock
    • !MCP Gateway GET and DELETE requests now return 405 because the gateway is stateless
30. Backend 1.24.2 Aug 15, 2026 · release notes

    Backend v1.24.2 adds configurable auto-create scopes for user API keys and improves JWKSJWKSA JSON Web Key Set, a published collection of cryptographic keys. Applications use it to find the public key needed to verify a token's signature. fetch resilience.

    • Adds user_api_key_auto_create_scopes organisation setting to control default data-plane scopes for auto-provisioned workspace-user API keys
    • Improves JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. authentication resilience in hybrid deployments by trying direct JWKS fetching before falling back to private link routing

    2 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The forwardHeaders parameter is removed from Portkey-managed PRO check schemas (portkey.moderateContent, portkey.language, portkey.pii, portkey.gibberish)
31. docs update Aug 14, 2026 · issue 045

    Portkey MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. gains rate limiting by request count or token consumption, plus content filters and approval workflows.

    • Adds rate limiting for MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. tool calls, throttleable by request count or token consumption, scoped to API key, server, or individual tool.
    • Adds content filters and approval workflows configurable per MCP server.

    2 details in Everything

32. docs update Aug 14, 2026 · issue 045

    Portkey adds MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. rate limiting, per-model custom host/headers, and an OCROCROptical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels. inference endpoint

    └──▷ TRY IT
    Throttle a specific MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. tool (send_message) on the Slack server to 100 calls per minute to protect downstream API quotas.
    copy

    $ curl -X POST "https://api.portkey.ai/v1/policies/rate-limits" \
      -H "x-portkey-api-key: <admin-api-key>" \
      -H "Content-Type: application/json" \
      -d '{
        "name": "Slack send_message limit",
        "description": "Throttle send_message calls on the Slack MCP server",
        "target": "mcp_tools",
        "limit": { "unit": "rpm", "value": 100 },
        "conditions": [
          { "key": "mcp_server", "value": "slack" },
          { "key": "mcp_tool", "value": "send_message" }
        ]
      }'

    Pin a custom upstream URL and tenant headers on a single Model Catalog entry so it routes to an internal regional endpoint without affecting other models in the same integration.
    copy

    $ curl -X PUT "https://api.portkey.ai/v1/integrations/YOUR_INTEGRATION_SLUG/models" \
      -H "x-portkey-api-key: <admin-api-key>" \
      -H "Content-Type: application/json" \
      -d '{
        "models": [{
          "slug": "my-custom-model-v1",
          "is_custom": true,
          "base_model_slug": "gpt-4o",
          "custom_host": "https://llm.internal.example.com/v1",
          "custom_headers": {
            "x-tenant-id": "acme-prod",
            "x-region": "us-east-1"
          }
        }]
      }'

    Cap total token consumption on the GitHub MCP server to 500,000 tokens per day to control costs across all users in a workspace.
    copy

    $ curl -X POST "https://api.portkey.ai/v1/policies/rate-limits" \
      -H "x-portkey-api-key: <admin-api-key>" \
      -H "Content-Type: application/json" \
      -d '{
        "name": "GitHub server token limit",
        "target": "mcp_tools",
        "limit": { "unit": "tpd", "value": 500000 },
        "conditions": [
          { "key": "mcp_server", "value": "github" }
        ]
      }'

    • Adds target: mcp_tools field to rate limit policies so MCP tool call traffic is governed independently from LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. traffic — scoped to workspace, server, tool, API key, or metadata.*.
    • Adds mcp_server and mcp_tool condition/group-by keys for rate limit policies, enabling per-server and per-tool throttling (e.g. cap the send_message tool to 100 rpm or the GitHub server to 500,000 tokens per day with unit: tpd).
    • Supports rpm, rph, rpd, and rpw (and token equivalents) as rate limit units for mcp_tools policies, with AND-logic multi-condition matching and an excludes modifier to carve out specific tools.
    • Creates MCP rate limit policies via POST https://api.portkey.ai/v1/policies/rate-limits with target set to mcp_tools; requires Gateway v2.18.0+ and Backend v1.24.0+.
    • Adds per-model custom_host and custom_headers fields to Model Catalog entries (set via PUT https://api.portkey.ai/v1/integrations/YOUR_INTEGRATION_SLUG/models under models[].configurations), letting different models in the same integration route to different upstreams or send different headers — model-level custom_headers override integration-level values.

    +1 more show less

    • Adds an OCR inference endpoint to extract text and structured content from PDFs and images.

    6 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !A rate limit policy's target field is immutable after creation — to switch between llm and mcp_tools, you must archive the existing policy and create a new one.
33. docs update Aug 14, 2026 · issue 045

    Portkey adds MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. rate limiting and a new OCROCROptical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels. inference endpoint for document text extraction.

    • Adds rate limiting for MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. tool call requests and token consumption via the MCP Gateway, documented under rate-limits.

    1 detail in Everything

34. Backend 1.24.1 Aug 12, 2026 · release notes

    Portkey Backend adds OAuth Client Credentials auth for OpenAI integrations and Azure workload identityworkload identityAn identity assigned to software, such as a service or container, rather than a person. It lets the workload authenticate and receive permissions without sharing a user's credentials. for self-hosted log stores.

    • Adds oauthClientCredentials as a valid value for the provider_auth_type field in integration configs to enable OAuth 2.0OAuth 2.0An open authorization framework (IETF RFC 6749) that lets a service grant limited API access to a third party without sharing credentials, giving cyber tools a standard way to authenticate against external platforms. Client Credentials grant for upstream providers like OpenAI
    • Adds support for AZURE_AUTH_MODE=workload environment variable to enable Azure Workload Identity authentication for log stores in self-hosted deployments

    2 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The forwardHeaders parameter is removed from 22 local-only guardrail check schemas (regex, model allowlist, word count, etc.) where it had no effect
35. Enterprise Gateway 2.18.0 Aug 10, 2026 · release notes

    Portkey Enterprise Gateway v2.18.0 adds native OCROCROptical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels. endpoint, Singulr guardrail, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. rate limits, and header forwarding.

    • Adds POST /v1/ocr endpoint to route OCR requests through gateway retries, fallbacks, load balancing, caching, logging, and pricing with support for Mistral AI and Azure AI Foundry
    • Adds Singulr partner guardrail plugin for request and response scanning via the standard guardrails workflow
    • Enables MCP tool call rate limits to throttle invocation volume per configured limits
    • Adds configurable guardrail header forwarding, allowing client headers like x-session-id and traceparent to be sent to upstream providers

    4 details in Everything

36. Backend 1.24.0 Aug 10, 2026 · release notes

    Backend 1.24.0 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. tool call rate limiting, forwardHeaders validation, A2AA2AAgent2Agent, an open protocol started by Google for AI agents from different vendors to discover each other, delegate tasks and exchange results. It complements MCP, which connects an agent to tools. analytics, and Redis auth for self-hosting

    • Adds target: mcp_tools to Rate Limit Policies for MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. enforcement via OAuth introspection and LIST filtering
    • Validates Guardrail forwardHeaders at create/update time and defines them in provider schemas for static export
    • Returns separate token, request, and error counters for LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code., MCP, and A2A traffic in Analytics chart APIs
    • Adds IAMIAMIdentity and Access Management — a framework of policies and services for controlling who can authenticate to systems and what resources they are authorized to use, letting cyber tools enforce least-privilege and audit access. authentication for AWS ElastiCache via AWS_REDIS_AUTH_MODE=iam and related env vars
    • Adds workload identityworkload identityAn identity assigned to software, such as a service or container, rather than a person. It lets the workload authenticate and receive permissions without sharing a user's credentials. support for Azure Cache for Redis via AZURE_REDIS_WORKLOAD_CLIENT_ID

    +1 more show less

    • Adds gcp-memory-store cache type with GCP workload identity auth via GCP_REDIS_AUTH_MODE

    6 details in Everything

37. Frontend 1.9.5 Aug 10, 2026 · release notes

    Portkey Frontend 1.9.5 adds Lightning AI provider support and model-level custom headers configuration.

38. updated seen Aug 6, 2026 · issue 037
    ⤢ ⤢
    └──▷ TRY IT
    Ship GCS logs from an AWS-hosted gateway without storing long-lived GCP credentials — set both WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. env vars and use the gcs_assume log storage flow.
    copy

    $ GCP_WIF_AUDIENCE=//iam.googleapis.com/projects/123/locations/global/workloadIdentityPools/my-pool/providers/my-provider \
    GCP_WIF_SERVICE_ACCOUNT_EMAIL=portkey-logger@my-project.iam.gserviceaccount.com \
    docker run portkey/gateway:latest

    • Adds Secret References so secret_mappings can point virtual keys and integrations at credentials stored in AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault — Portkey fetches values at runtime instead of storing keys directly.
    • Adds weekly (rpw) budget and usage windows so rate-limit caps align with weekly spend planning, not only per-minute or monthly aggregates.
    • Adds endpoint-scoped rate limits so chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes can each carry independent limits instead of one global rule.
    • Adds GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL environment variables to authenticate GCS log writes through GCP Workload Identity Federation (gcs_assume flow) when the gateway runs in AWS, eliminating long-lived GCP keys in AWS config.
    • Enables analytics for archived workspaces and accepts workspace slugs (alongside IDs) in saved filters, keeping reporting stable as workspaces are wound down or renamed.

    +2 more show less

    • Integrates Zscaler AI Guard as a guardrail partner via beforeRequestHook and afterRequestHook, requiring a policyId and supporting a configurable timeout (default 10000 ms).
    • Maps Portkey metadata to Vertex AI resource labels; adds enterprise search grounding via enterpriseWebSearch / enterprise_web_search; enables AWS workloads to reach Vertex AI using GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL.
    └──▷ BREAKING ON UPGRADE

    • !The gateway no longer forwards the x-portkey-forward-headxers header, preventing header-forwarding loops — any setup relying on that header being passed through will stop working.
39. updated seen Aug 5, 2026 · issue 036
    ⤢ ⤢
    └──▷ TRY IT
    Route GCS log writes through GCP Workload Identity Federation when the Portkey gateway runs in AWS, avoiding static GCP service-account keys.
    copy

    $ GCP_WIF_AUDIENCE=//iam.googleapis.com/projects/123456/locations/global/workloadIdentityPools/my-pool/providers/my-provider \
    GCP_WIF_SERVICE_ACCOUNT_EMAIL=portkey-logs@my-project.iam.gserviceaccount.com \
    docker run portkey/gateway:latest

    Apply Zscaler AI Guard detection policies to all LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. requests and responses with a custom evaluation timeout.
    json copy

    {
      "guardrails": [
        {
          "provider": "zscaler",
          "hooks": ["beforeRequestHook", "afterRequestHook"],
          "policyId": "your-zscaler-policy-id",
          "timeout": 8000
        }
      ]
    }

    • Adds Secret References so credentials stored in AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault are fetched at runtime via secret_mappings, keeping plaintext keys out of Portkey config.
    • Adds weekly budget/usage windows (rpw) and endpoint-scoped rate limits, letting teams cap spend per API surface (chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., admin routes) independently.
    • Adds Zscaler AI Guard guardrail integration applying Zscaler Detection Policies via beforeRequestHook / afterRequestHook, with required policyId and optional timeout (default 10 000 ms).
    • Adds required metadata key–value guardrails that block requests missing or carrying invalid required metadata fields before they reach the model.
    • Adds tools, tool_choice, and stream_options for deepseek-chat; maps reasoning_effort to thinking mode and surfaces reasoning_content in streams for deepseek-reasoner.

    +2 more show less

    • Adds Vertex AI enterprise search grounding via enterpriseWebSearch / enterprise_web_search (with separate cost attribution) and AWS-to-Vertex connectivity using GCP_WIF_AUDIENCE / GCP_WIF_SERVICE_ACCOUNT_EMAIL.
    • Adds Vertex AI resource label support by mapping Portkey metadata to Vertex resource labels.
40. Backend 1.23.0 Aug 5, 2026 · release notes

    Portkey Backend 1.23.0 introduces hardened SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. checks for gateway URLs and a new DB index.

    • Adds hardened SSRF checks for prompt playground and proxy calls by blocking IMDS, metadata endpoints, and private DNS-rebind targets for admin-supplied gateway URLs
    • Auto-adds env-configured gateway hostnames to the trusted host set
    • Adds index on api_keys.last_updated_at for improved query performance

    3 details in Everything

41. updated seen Aug 4, 2026 · issue 035

    Portkey March 2026 adds vault-backed Secret References, weekly/endpoint-scoped rate limits, cross-cloud GCS log delivery, and four new guardrail integrations.

    ⤢ ⤢
    └──▷ USE IT
    Keep LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. provider credentials out of Portkey by pointing a virtual key at a secret already stored in AWS Secrets Manager.
    json copy

    {
      "virtual_key": "my-openai-key",
      "secret_mappings": {
        "api_key": "arn:aws:secretsmanager:us-east-1:123456789012:secret:prod/openai-api-key"
      }
    }

    Route GCS log writes through GCP Workload Identity Federation when the Portkey gateway runs in AWS, avoiding long-lived GCP credentials in your AWS environment.
    copy

    $ GCP_WIF_AUDIENCE=//iam.googleapis.com/projects/123456/locations/global/workloadIdentityPools/my-pool/providers/my-provider \
    GCP_WIF_SERVICE_ACCOUNT_EMAIL=portkey-log-writer@my-project.iam.gserviceaccount.com \
    docker run portkey/gateway ...

    • Adds Secret References so credentials stored in AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault can be mapped to integrations and virtual keys via secret_mappings, with Portkey fetching values at runtime instead of storing them directly.
    • Adds rpw (requests-per-week) budget and usage windows so rate limits can align with weekly planning cycles, alongside endpoint-scoped limits that let chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes carry independent caps.
    • Adds GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL environment variables to enable GCP Workload Identity Federation (gcs_assume style) for GCS log delivery when the gateway runs in AWS, eliminating static cross-cloud credentials.
    • Adds Zscaler AI Guard guardrail integration using beforeRequestHook and afterRequestHook hooks, requiring a policyId and supporting a configurable timeout (default 10000 ms) to apply Zscaler Detection Policies to LLM inputs and outputs.
    • Adds Akto Agentic Security as a guardrails partner to scan LLM inputs and outputs for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage via hooks with a configurable timeout (default 5000 ms).

    +9 more show less

    • Adds customHost configuration to the Bedrock guardrail plugin so evaluations can target private or regional Bedrock-compatible endpoints instead of the default public Bedrock URLs.
    • Adds required metadata key–value guardrails that block requests at the gateway before they reach the model if required metadata fields are missing or invalid.
    • Extends analytics so organization admins and owners can include archived workspaces in graphs and summaries; saved filters now accept workspace slugs alongside IDs.
    • Adds DeepInfra provider support with tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. via tools, tool_choice, and parallel_tool_calls, plus completions and embeddings endpoints.
    • Adds stream_options support for DeepSeek deepseek-chat, and maps reasoning_effort to thinking mode for deepseek-reasoner, returning reasoning_content in streams.
    • Extends Bedrock batch inference to support embeddings in addition to chat completions.
    • Adds Vertex AI enterprise web search grounding via enterpriseWebSearch / enterprise_web_search with separate cost attribution, Portkey metadata mapping to Vertex resource labels, and AWS-to-Vertex access via GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL.
    • Adds Azure AI Foundry rerank support for Cohere rerank models (e.g. cohere.Cohere-rerank-v4.0-pro), with the gateway stripping the cohere. prefix for the provider.
    • OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. GenAI semantic spans now follow semconv 1.40.0 for inference and embeddings, with OTELOTELAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. exporter support for guardrail flows and custom resource attributes.

    14 details in Everything

42. Enterprise Gateway 2.17.0 Aug 3, 2026 · release notes

    Enterprise Gateway 2.17.0 adds Databricks Responses API support, new Anthropic beta parameters, and a beta flag for Responses API 2026-07-30

    • Databricks provider now supports the Responses API, enabling OpenAI-compatible Responses format for Databricks-served models
    • New Anthropic beta parameters can now be passed through to the Anthropic provider: inference_geo, diagnostics, fallbacks, and context_management
    • Beta flag use-responses-api-2026-07-30 enables the latest Responses API schema version via the x-portkey-beta header

    3 details in Everything

43. Backend 1.22.0 Aug 3, 2026 · release notes

    Portkey Backend 1.22.0 adds organisation guardrail defaults, workspace exclusions via API, and new OAuth scopes.

    • Adds GET /v1/workspace-exclusions/input-guardrails and /v1/workspace-exclusions/output-guardrails to exempt workspaces from org-level defaults.
    • Adds PUT /v1/admin/organisation/defaults to set organisation-wide input_guardrails and output_guardrails.
    • New API scopes organisation_settings.*, organisation_guardrails.*, and organisation_exclusions.* for fine-grained access control.
    • OIDCOIDCOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user. settings now accept token_endpoint_auth_method and jwks key set, enabling private_key_jwt authentication.
    • Model Catalog GET provider details accepts include_configurations=true to retrieve system auth configs with masked secrets.

    5 details in Everything

44. docs update seen Jul 30, 2026 · issue 030

    Portkey Enterprise adds configurable custom Gateway URLs for self-hosted AI and MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. gateway deployments.

    └──▷ HOW TO FIND IT
    Point the Portkey dashboard at your self-hosted AI gateway so all generated code snippets and playground completions route through your own infrastructure.
    📍In the dashboard, go to Settings › Organisation › General, scroll to the Gateway URLs section, enter your base URL in the Gateway URL field (e.g. https://gateway.yourcompany.com), then click outside the field to save.
    Register a self-hosted MCP gateway so MCP server endpoint URLs (e.g. https://mcp.yourcompany.com/my-server/mcp) are constructed from your own domain.
    📍In the dashboard, go to Settings › Organisation › General, scroll to the Gateway URLs section, enter your MCP gateway base URL in the MCP Gateway URL field (e.g. https://mcp.yourcompany.com), then click outside the field to save.
    • Adds organisation-level Gateway URL setting so self-hosted AI gateway deployments replace Portkey's cloud endpoint (https://api.portkey.ai/v1) in code snippets, Model Catalog, and playground completions.
    • Adds organisation-level MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. URL setting to point MCP server endpoint construction at a self-hosted MCP gateway instead of https://mcp.portkey.ai.
    • Dashboard auto-normalises trailing /v1 on the Gateway URL field, saving automatically on focus-out with no separate save button.

    3 details in Everything

45. updated seen Jul 30, 2026 · issue 030

    Portkey adds vault-backed secret references, weekly/endpoint-scoped rate limits, two new guardrail integrations, and broader provider coverage.

    └──▷ USE IT
    Apply Zscaler Detection Policies to all LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. traffic by wiring Zscaler AI Guard into your Portkey config as a guardrail hook.
    json copy

    {
      "guardrails": [
        {
          "type": "zscaler",
          "hooks": ["beforeRequestHook", "afterRequestHook"],
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      ]
    }

    Set a weekly spend cap scoped only to the embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. endpoint so it doesn't share budget with chat completions.
    📍In the Portkey console, go to Budget & Rate Limit Policies, create a new policy, set the window to rpw (weekly), set the endpoint scope to embeddings, and enter your desired token or cost limit.
    • Introduces weekly rate/budget windows (rpw) and endpoint-scoped rate limits, letting teams cap spend per API surface (chat completions, embeddings, admin routes, etc.) independently.

    1 detail in Everything

46. docs update seen Jul 28, 2026 · issue 028

    Portkey forwards request metadata as AWS STSAWS STSAWS Security Token Service, an Amazon Web service that issues short-lived credentials for IAM roles, letting tools authenticate to AWS APIs without storing long-term secrets. session tags, enabling per-app Bedrock cost attribution in Cost Explorer and CloudTrailCloudTrailAWS service that records API calls and account activity across an AWS environment, giving cyber tools an audit log to detect anomalous or malicious actions..

    └──▷ TRY IT
    Attribute Bedrock LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. costs to a specific app, team, and environment in AWS Cost Explorer by tagging requests at the gateway level.
    copy

    $ curl https://api.portkey.ai/v1/chat/completions \
      -H "x-portkey-provider: $PROVIDER_SLUG" \
      -H 'x-portkey-metadata: {"app": "chatbot", "team": "platform", "env": "prod"}' \
      -d '{ "model": "anthropic.claude-3-5-sonnet-20240620-v1:0", "messages": [{"role": "user", "content": "Hello"}] }'

    • Supports AWS STSSTSAWS Security Token Service, an Amazon Web Services API that issues short-lived credentials for IAM roles, letting cyber tools authenticate to AWS without storing long-term secrets. session tags for Bedrock: metadata sent via x-portkey-metadata is forwarded as STS session tags, enabling spend segregation in AWS Cost Explorer and CloudTrail without separate IAMIAMIdentity and Access Management — a framework of policies and services for controlling who can authenticate to systems and what resources they are authorized to use, letting cyber tools enforce least-privilege and audit access. roles per application.
    • Clarifies that model conditions and group_by keys use the request model name (e.g. gpt-4.1), not the provider's response model name (e.g. gpt-4.1-2025-06-01), which matters for Azure OpenAI and similar providers.

    2 details in Everything

47. docs update seen Jul 28, 2026 · docs issue 004

    Portkey forwards request metadata as AWS STSAWS STSAWS Security Token Service, an Amazon Web service that issues short-lived credentials for IAM roles, letting tools authenticate to AWS APIs without storing long-term secrets. session tags, enabling per-app Bedrock cost attribution in Cost Explorer and CloudTrailCloudTrailAWS service that records API calls and account activity across an AWS environment, giving cyber tools an audit log to detect anomalous or malicious actions..

    └──▷ TRY IT
    Attribute Bedrock LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. costs to a specific app, team, and environment in AWS Cost Explorer by tagging requests at the gateway level.
    copy

    $ curl https://api.portkey.ai/v1/chat/completions \
      -H "x-portkey-provider: $PROVIDER_SLUG" \
      -H 'x-portkey-metadata: {"app": "chatbot", "team": "platform", "env": "prod"}' \
      -d '{ "model": "anthropic.claude-3-5-sonnet-20240620-v1:0", "messages": [{"role": "user", "content": "Hello"}] }'

    • Supports AWS STSSTSAWS Security Token Service, an Amazon Web Services API that issues short-lived credentials for IAM roles, letting cyber tools authenticate to AWS without storing long-term secrets. session tags for Bedrock: metadata sent via x-portkey-metadata is forwarded as STS session tags, enabling spend segregation in AWS Cost Explorer and CloudTrail without separate IAMIAMIdentity and Access Management — a framework of policies and services for controlling who can authenticate to systems and what resources they are authorized to use, letting cyber tools enforce least-privilege and audit access. roles per application.
    • Clarifies that model conditions and group_by keys use the request model name (e.g. gpt-4.1), not the provider's response model name (e.g. gpt-4.1-2025-06-01), which matters for Azure OpenAI and similar providers.

    2 details in Everything

48. docs update seen Jul 28, 2026 · issue 028

    Portkey adds support for configuring custom gateway endpoints for organizations.

49. docs update seen Jul 28, 2026 · docs issue 004

    Portkey adds support for configuring custom gateway endpoints for organizations.

50. updated seen Jul 28, 2026 · issue 028

    Portkey March 2026: vault-backed credentials, weekly rate limits, cross-cloud GCS logging, and four new guardrail integrations.

    └──▷ USE IT
    Apply Zscaler AI Guard detection policies to all LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. inputs and outputs by wiring in your existing policyId.
    json copy

    {
      "guardrails": [
        {
          "type": "zscaler",
          "hooks": ["beforeRequestHook", "afterRequestHook"],
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      ]
    }

    • Adds Secret References to pull credentials from AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault at runtime via secret_mappings, keeping secrets out of Portkey config.
    • Supports weekly budget and usage windows (rpw) for rate limits, enabling caps aligned to weekly planning and review cycles.
    • Adds endpoint-scoped rate limits so chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes can carry independent limits instead of a single global rule.
    • Enables GCP Workload Identity Federation log delivery from AWS-hosted gateways to GCS via GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL, eliminating long-lived cross-cloud keys.
    • Allows organization admins to include archived workspaces in analytics graphs and summaries; saved filters now accept workspace slugs alongside IDs.

    +11 more show less

    • Integrates Zscaler AI Guard as a guardrail partner, applying Zscaler Detection Policies to LLM inputs and outputs via beforeRequestHook and afterRequestHook with a required policyId.
    • Integrates Akto Agentic Security as a guardrail partner for scanning LLM inputs and outputs for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage.
    • Adds customHost to the Bedrock Guardrails plugin, routing guardrail evaluation to private or regional Bedrock-compatible endpoints.
    • Adds required metadata key–value guardrails that block requests missing specified metadata fields before they reach the model.
    • Adds DeepInfra provider support with tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. (tools, tool_choice, parallel_tool_calls), completions, and embeddings endpoints.
    • Expands DeepSeek support: deepseek-chat gains tools, tool_choice, and stream_options; deepseek-reasoner maps reasoning_effort to thinking mode and returns reasoning_content in streams.
    • Extends Bedrock batch inference to support embeddings jobs alongside existing chat completions batch patterns.
    • Maps Portkey metadata to Vertex AI resource labels and adds enterprise search grounding via enterpriseWebSearch / enterprise_web_search with separate cost attribution.
    • Enables AWS workloads to reach Vertex AI through AWS–GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. using GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL.
    • Adds Azure AI Foundry rerank support for Cohere rerank models (e.g., cohere.Cohere-rerank-v4.0-pro), with the gateway stripping the cohere. prefix for the provider.
    • Updates OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. GenAI semantic spans to semconv 1.40.0 for inference and embeddings, with OTELOTELAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. exporter support for guardrail flows and custom resource attributes.

    16 details in Everything

51. Enterprise Gateway 2.16.0 Jul 27, 2026 · release notes

    Portkey Enterprise Gateway 2.16 adds Responses API routing, Lightning AI provider, Akto guardrails, and MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. user attribution.

    • Routes Anthropic Messages API requests to OpenAI Responses API models (e.g. codex) with automatic transformation
    • Adds Lightning AI provider integration for chat completions
    • Enables Akto guardrail scanning for both request and response payloads
    • Attributes MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. log entries to specific users via OAuth claims or API key name

    4 details in Everything

52. Frontend 1.9.4 Jul 27, 2026 · release notes

    Portkey Frontend 1.9.4 adds SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. client ID login, SSO auto-provisioning toggle, Soft Deny guardrails mode, and MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. transparency UI.

    • Adds SSO Client ID login support with improved redirect flow when organisation_id is present
    • Adds UI toggle for SSO auto user provisioning
    • Adds Soft Deny (200) toggle for guardrails, allowing denied requests to return a 200 response instead of a hard block
    • MCP consent page now displays the redirect URI for transparency
    • Adds org and workspace details on the MCP consent no access page

    +1 more show less

    • Adds Databricks to the provider list in the dashboard

    6 details in Everything

53. docs update seen Jul 24, 2026 · issue 024

    Portkey adds Prisma AIRS agentic scan controls, soft-deny guardrails, provider-backed model listing, and Continue.dev integration

    └──▷ USE IT
    Reduce Prisma AIRS false positives when using agentic clients (Claude Code, Cursor, Cline) by scanning only the last user message and stripping scaffolding wrappers.
    json copy

    {
      "scan_scope": "last_user_message",
      "strip_scaffolding": true
    }

    Keep an agentic session alive when a guardrail fires by returning a soft deny as HTTP 200 instead of HTTP 446.
    json copy

    {
      "soft_deny_200": true
    }

    • Adds scan_scope and strip_scaffolding parameters to the Palo Alto Networks Prisma AIRS guardrail, enabling fine-grained control over which messages are scanned and reducing false positives for agentic clients like Claude Code, Cursor, and Cline.
    • Adds soft_deny_200 flag on guardrail checks: denials return HTTP 200 with the error as a chat completion response instead of HTTP 446, keeping sessions alive for clients that treat 4xx as fatal.
    • GET /v1/models now routes through the gateway's provider-backed model listing handler, calling the upstream provider and returning a normalized model list with response transforms applied.
    • Anthropic cost attribution now differentiates batch (50% discount), fast, and standard pricing tiers for multidimensional cost tracking.
    • xAI server-side tool invocation costs (file search, web search) are now tracked for cost attribution.

    +1 more show less

    • New integration guide for Continue.dev: adds enterprise-grade observability, cost tracking, and governance to the open-source AI coding assistant via Portkey as an OpenAI-compatible endpoint.

    6 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Configs exceeding the new fan-out limits (100 root-level targets, 1,000 for conditional strategy, 50 nested targets per root target) will be rejected at request time; override with MAX_ROOT_CONFIG_TARGETS, MAX_NESTED_CONFIG_TARGETS, or MAX_CONDITIONAL_ROOT_CONFIG_TARGETS environment variables before upgrading.
54. docs update seen Jul 24, 2026 · docs issue 004

    Portkey adds Prisma AIRS agentic scan controls, soft-deny guardrails, provider-backed model listing, and Continue.dev integration

    └──▷ USE IT
    Reduce Prisma AIRS false positives when using agentic clients (Claude Code, Cursor, Cline) by scanning only the last user message and stripping scaffolding wrappers.
    json copy

    {
      "scan_scope": "last_user_message",
      "strip_scaffolding": true
    }

    Keep an agentic session alive when a guardrail fires by returning a soft deny as HTTP 200 instead of HTTP 446.
    json copy

    {
      "soft_deny_200": true
    }

    • Adds scan_scope and strip_scaffolding parameters to the Palo Alto Networks Prisma AIRS guardrail, enabling fine-grained control over which messages are scanned and reducing false positives for agentic clients like Claude Code, Cursor, and Cline.
    • Adds soft_deny_200 flag on guardrail checks: denials return HTTP 200 with the error as a chat completion response instead of HTTP 446, keeping sessions alive for clients that treat 4xx as fatal.
    • GET /v1/models now routes through the gateway's provider-backed model listing handler, calling the upstream provider and returning a normalized model list with response transforms applied.
    • Anthropic cost attribution now differentiates batch (50% discount), fast, and standard pricing tiers for multidimensional cost tracking.
    • xAI server-side tool invocation costs (file search, web search) are now tracked for cost attribution.

    +1 more show less

    • New integration guide for Continue.dev: adds enterprise-grade observability, cost tracking, and governance to the open-source AI coding assistant via Portkey as an OpenAI-compatible endpoint.

    6 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Configs exceeding the new fan-out limits (100 root-level targets, 1,000 for conditional strategy, 50 nested targets per root target) will be rejected at request time; override with MAX_ROOT_CONFIG_TARGETS, MAX_NESTED_CONFIG_TARGETS, or MAX_CONDITIONAL_ROOT_CONFIG_TARGETS environment variables before upgrading.
55. docs update seen Jul 24, 2026 · issue 024

    Portkey adds MCP ServerMCP ServerA process that exposes tools, resources, and prompts over Anthropic's Model Context Protocol, letting an AI agent call external APIs or local services through a standardised interface. management, SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. workspace mappings, rate/usage limit policies, and expanded Admin APIs for guardrails, models, and prompts.

    • New Admin APIAdmin APIA dedicated HTTP interface that exposes management and configuration operations separately from the data plane, letting cyber tools automate provisioning, policy changes, or monitoring without touching the main user-facing service. endpoints to create, list, and test MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Server connections for workspace instances.
    • New Admin API endpoints to manage SCIM group-to-workspace mappings (create, list, delete) for enterprise identity provisioning.
    • New Admin API endpoints to create and list Rate Limits Policies and Usage Limits Policies, with entity tracking for usage limits.
    • New Admin API endpoint to rotate existing API keys and retrieve a new key.
    • New Admin API endpoints to list and retrieve individual Guardrail configurations.

    +8 more show less

    • New Admin API endpoints for custom model management: delete custom models, list model access, and update model access and pricing configuration.
    • New Admin API endpoints to list and update workspace access permissions and usage limits per integration.
    • New Admin API endpoints to create and list Prompt Collections, update prompt metadata, and update prompt version metadata.
    • New Skills Registry for creating versioned, team-shared skills usable by AI coding agents via Portkey.
    • New guide for converting local STDIO MCP servers to remote Streamable HTTP MCPStreamable HTTP MCPA transport variant of the Model Context Protocol (MCP) that carries tool-call messages over HTTP with optional streaming, letting cyber tools expose AI-callable functions without maintaining a persistent connection. servers via Portkey's MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once..
    • New integration support for AWS AgentCore, Agno AI, enabling Portkey-powered agents inside Amazon Bedrock and Agno.
    • New use-case guide for using private MCP Servers with the Responses API via client-side tool handling.
    • New C# (.NET) SDK support for integrating Portkey in .NET applications.

    13 details in Everything

56. docs update seen Jul 24, 2026 · docs issue 004

    Portkey adds MCP ServerMCP ServerA process that exposes tools, resources, and prompts over Anthropic's Model Context Protocol, letting an AI agent call external APIs or local services through a standardised interface. management, SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. workspace mappings, rate/usage limit policies, and expanded Admin APIs for guardrails, models, and prompts.

    • New Admin APIAdmin APIA dedicated HTTP interface that exposes management and configuration operations separately from the data plane, letting cyber tools automate provisioning, policy changes, or monitoring without touching the main user-facing service. endpoints to create, list, and test MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Server connections for workspace instances.
    • New Admin API endpoints to manage SCIM group-to-workspace mappings (create, list, delete) for enterprise identity provisioning.
    • New Admin API endpoints to create and list Rate Limits Policies and Usage Limits Policies, with entity tracking for usage limits.
    • New Admin API endpoint to rotate existing API keys and retrieve a new key.
    • New Admin API endpoints to list and retrieve individual Guardrail configurations.

    +8 more show less

    • New Admin API endpoints for custom model management: delete custom models, list model access, and update model access and pricing configuration.
    • New Admin API endpoints to list and update workspace access permissions and usage limits per integration.
    • New Admin API endpoints to create and list Prompt Collections, update prompt metadata, and update prompt version metadata.
    • New Skills Registry for creating versioned, team-shared skills usable by AI coding agents via Portkey.
    • New guide for converting local STDIO MCP servers to remote Streamable HTTP MCPStreamable HTTP MCPA transport variant of the Model Context Protocol (MCP) that carries tool-call messages over HTTP with optional streaming, letting cyber tools expose AI-callable functions without maintaining a persistent connection. servers via Portkey's MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once..
    • New integration support for AWS AgentCore, Agno AI, enabling Portkey-powered agents inside Amazon Bedrock and Agno.
    • New use-case guide for using private MCP Servers with the Responses API via client-side tool handling.
    • New C# (.NET) SDK support for integrating Portkey in .NET applications.

    13 details in Everything

57. updated seen Jul 24, 2026 · issue 024

    Portkey adds vault-backed Secret References, weekly rate limits, cross-cloud GCS log delivery, and four new guardrail integrations.

    └──▷ USE IT
    Apply Zscaler Detection Policies to all LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. traffic through Portkey so your existing Zscaler policy class governs both inputs and outputs.
    json copy

    Add Zscaler AI Guard as a guardrail with your policyId:
    
    {
      "guardrails": {
        "beforeRequestHook": {
          "provider": "zscaler",
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        },
        "afterRequestHook": {
          "provider": "zscaler",
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      }
    }

    • Adds Secret References so virtual keys and integrations can pull credentials at runtime from AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault via secret_mappings, keeping secrets out of Portkey config.
    • Supports weekly budget and usage windows (rpw) for rate-limit policies so spend caps align with weekly planning cycles.
    • Adds endpoint-scoped rate limits so chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes can each carry independent limits instead of a single global cap.
    • Enables GCS log storage from AWS-hosted gateways via GCP Workload Identity Federation (GCP_WIF_AUDIENCE, GCP_WIF_SERVICE_ACCOUNT_EMAIL), eliminating long-lived cross-cloud credentials.
    • Lets organization admins include archived workspaces in analytics graphs and filters, with workspace slugs now accepted alongside IDs in saved filters.

    +11 more show less

    • Integrates Zscaler AI Guard as a guardrail partner, applying Zscaler Detection Policies to LLM inputs and outputs via beforeRequestHook/afterRequestHook with a required policyId.
    • Integrates Akto Agentic Security as a guardrail partner to scan LLM traffic for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage.
    • Adds customHost support for the Bedrock Guardrails plugin so evaluations route to private or regional Bedrock-compatible endpoints.
    • Adds required metadata key–value guardrails that block requests missing specified fields before they reach the model.
    • Adds DeepInfra provider support with tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. (tools, tool_choice, parallel_tool_calls), chat completions, and embeddings.
    • Extends DeepSeek support: deepseek-chat gains tools, tool_choice, and stream_options; deepseek-reasoner maps reasoning_effort to thinking mode and surfaces reasoning_content in streams.
    • Extends Bedrock batch inference to cover embeddings in addition to chat completions.
    • Maps Portkey metadata to Vertex AI resource labels and adds enterprise web search grounding via enterpriseWebSearch/enterprise_web_search.
    • Enables AWS workloads to reach Vertex AI through AWS–GCP Workload Identity Federation (GCP_WIF_AUDIENCE, GCP_WIF_SERVICE_ACCOUNT_EMAIL).
    • Adds Azure AI Foundry rerank support for Cohere rerank models (e.g. cohere.Cohere-rerank-v4.0-pro), with the gateway stripping the cohere. prefix for the provider.
    • Updates OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. GenAI semantic spans to semconv 1.40.0, adds OTELOTELAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. exporter support for guardrail flows, and supports custom resource attributes for downstream APMAPMApplication Performance Monitoring — a category of software that collects metrics, traces, and logs from running applications so developers can detect slowdowns and diagnose their root cause. and tracing.

    16 details in Everything

58. docs update seen Jul 21, 2026 · issue 021

    Portkey's GET /v1/models now supports provider-backed proxying and a new header to force Model Catalog results.

    └──▷ TRY IT
    Show end users only your workspace's approved Model Catalog models, even when the request carries provider auth for downstream routing.
    copy

    $ curl https://api.portkey.ai/v1/models \
      -H "x-portkey-api-key: <your-portkey-key>" \
      -H "x-portkey-provider: @anthropic-prod" \
      -H "x-portkey-fetch-integrated-models: true"

    • Adds provider-backed mode to GET /v1/models: pass x-portkey-provider: @<provider-slug> or a config with provider/targets to proxy the request to the upstream provider's /v1/models and receive a response normalized to OpenAI models list format.
    • Adds x-portkey-fetch-integrated-models: true header (and fetch_integrated_models: true config field) to force GET /v1/models to return your workspace's Model Catalog models even when a provider header or config is present.
    • Model Catalog results are automatically filtered to only the provider slugs referenced in your routing config when fetch_integrated_models is active; passthrough: true targets return the full unfiltered list.

    3 details in Everything

59. docs update seen Jul 21, 2026 · issue 021

    Portkey adds a Models API endpoint to retrieve the list of available models.

    └──▷ TRY IT
    Query the Models endpoint to discover which models are available to your account before routing requests.
    copy

    $ curl https://api.portkey.ai/v1/models -H "x-portkey-api-key: <your-api-key>"

60. updated seen Jul 21, 2026 · issue 021

    Portkey adds vault-backed credentials, weekly rate limits, cross-cloud GCS logging, and four new guardrail integrations in March 2026.

    └──▷ HOW TO FIND IT
    Pull API keys from your own vault at request time instead of storing them in Portkey — critical for orgs with strict secrets governance.
    📍In the Portkey console, go to your Virtual Key or Integration settings and add a secret_mappings block pointing to your AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault path. Portkey will resolve the credential at runtime.
    Cap a team's weekly LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. spend and keep chat completions under a separate limit from embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. calls.
    📍In the Portkey console, open Budget & Rate Limit Policies, set a rpw (requests-per-week) budget window for the relevant virtual key or config, then add an endpoint-scoped limit targeting the chat completions route independently from embeddings.
    Route Bedrock guardrail evaluation to a private regional endpoint instead of the default public Bedrock URL.
    📍In the Portkey console, open the Bedrock guardrail plugin configuration and set customHost to your private or regional Bedrock-compatible endpoint URL.
    • Introduces weekly rate and budget windows (rpw) so spend caps align with team planning cycles, not just per-minute or monthly aggregates.
    • Adds endpoint-scoped rate limits so chat completions, embeddings, and admin routes can each carry independent limits instead of a single global rule.
    • Enables GCP Workload Identity Federation log storage (gcs_assume) from AWS-hosted gateways via GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL, eliminating static GCP keys in AWS.
    • Allows organization admins to include archived workspaces in analytics graphs and filters; saved filters now also accept workspace slugs alongside IDs.
    • Integrates Zscaler AI Guard as a guardrails partner, applying Zscaler Detection Policies to LLM inputs and outputs via beforeRequestHook/afterRequestHook with a required policyId.

    +6 more show less

    • Integrates Akto Agentic Security as a guardrails partner to scan LLM traffic for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage via configurable hooks and timeout.
    • Adds required metadata key–value guardrails that block requests missing or carrying invalid metadata fields before they reach the model.
    • Supports DeepInfra tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. (tools, tool_choice, parallel_tool_calls) plus completions and embeddings endpoints.
    • Adds DeepSeek deepseek-chat tool calling (tools, tool_choice, stream_options) and deepseek-reasoner reasoning effort mapping with reasoning_content in streams.
    • Extends Bedrock batch inference to support embeddings in addition to chat completions.
    • Adds Azure AI Foundry rerank support for Cohere rerank models (e.g., cohere.Cohere-rerank-v4.0-pro), with the gateway stripping the cohere. prefix for the provider.

    11 details in Everything

61. docs update seen Jul 20, 2026 · issue 020

    Portkey adds hybrid deployment architecture option.

    • Supports hybrid deployment architecture, enabling mixed on-premises and cloud configurations.

    1 detail in Everything

62. updated seen Jul 20, 2026 · issue 020
    └──▷ USE IT
    Gate every agentic request through Zscaler AI Guard so your existing Zscaler Detection Policies apply to LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. inputs and outputs.
    json copy

    In your Portkey config, add the Zscaler AI Guard guardrail with your policyId:
    {
      "guardrails": {
        "beforeRequestHook": {
          "provider": "zscaler",
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        },
        "afterRequestHook": {
          "provider": "zscaler",
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      }
    }

    Route GCP Cloud Storage log writes from an AWS-hosted gateway through Workload Identity Federation to avoid storing long-lived GCP service account keys in AWS.
    copy

    $ Set these environment variables on your gateway deployment:
    GCP_WIF_AUDIENCE=<your-wif-audience>
    GCP_WIF_SERVICE_ACCOUNT_EMAIL=<your-sa>@<project>.iam.gserviceaccount.com

    • Adds Secret References so Portkey fetches credentials at runtime from AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault via secret_mappings, keeping keys out of Portkey config entirely.
    • Enables GCP Workload Identity Federation log delivery from AWS-hosted gateways to Google Cloud Storage via GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL, eliminating long-lived GCP keys in AWS.
    • Adds customHost support on the Bedrock Guardrails plugin so guardrail evaluation hits private or regional Bedrock-compatible endpoints instead of default public URLs.
    • Adds tool_choice, stream_options, and streaming usage reporting to DeepSeek deepseek-chat; maps reasoning_effort to thinking mode and surfaces reasoning_content in streams for deepseek-reasoner.
    • Maps Portkey metadata to Vertex AI resource labels and adds enterprise search grounding via enterpriseWebSearch / enterprise_web_search with separate cost attribution.

    +2 more show less

    • Enables AWS workloads to reach Vertex AI through AWS–GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. using GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL.
    • Updates OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. GenAI semantic spans to semconv 1.40.0, adds OTELOTELAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. exporter support for guardrail flows, and supports custom resource attributes for downstream APMAPMApplication Performance Monitoring — a category of software that collects metrics, traces, and logs from running applications so developers can detect slowdowns and diagnose their root cause. standardization.
63. updated seen Jul 17, 2026 · issue 018

    Portkey adds vault-backed credentials, weekly rate limits, cross-cloud GCS logging, and four new guardrail integrations in March 2026.

    └──▷ HOW TO FIND IT
    Pull LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. provider credentials from your existing vault at runtime instead of storing keys in Portkey directly.
    📍In the Portkey console, go to Settings › Integrations (or Virtual Keys) and add a secret_mappings entry pointing to your AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault secret ARN/path. Portkey fetches the value at request time.
    Route Zscaler AI Guard policy checks across all LLM traffic to enforce your org's existing detection policies on inputs and outputs.
    json copy

    {
      "guardrails": [
        {
          "type": "zscaler",
          "beforeRequestHook": true,
          "afterRequestHook": true,
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      ]
    }

    Ship gateway logs from AWS to GCS using Workload Identity Federation — no long-lived GCP keys required.
    copy

    $ GCP_WIF_AUDIENCE=<your-wif-audience>
    GCP_WIF_SERVICE_ACCOUNT_EMAIL=<your-service-account>@<project>.iam.gserviceaccount.com
    # Set log storage mode to gcs_assume in your gateway config

    • New Secret References let you point Portkey at credentials stored in AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault via secret_mappings, so keys are fetched at runtime from vaults you already control.
    • Adds weekly rate and budget windows (rpw) so spend caps align with weekly planning cycles rather than only per-minute or monthly aggregates.
    • Adds endpoint-scoped rate limits so chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes can each carry independent limits instead of a single global rule.
    • GCP Workload Identity Federation log delivery: configure GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL so an AWS-hosted gateway writes logs to GCS without long-lived GCP keys (gcs_assume style).
    • Organization admins can now include archived workspaces in analytics graphs and filters; saved filters also accept workspace slugs alongside IDs.

    +9 more show less

    • New Zscaler AI Guard integration applies Zscaler Detection Policies to LLM inputs and outputs via beforeRequestHook / afterRequestHook with a required policyId.
    • New Akto Agentic Security guardrail scans LLM inputs and outputs for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage with configurable hook and timeout.
    • Bedrock Guardrails plugin now accepts a customHost field to route guardrail evaluation to private or regional Bedrock-compatible endpoints.
    • New required metadata key–value guardrails block requests missing specified metadata fields before they reach the model.
    • DeepInfra provider adds tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. (tools, tool_choice, parallel_tool_calls) plus completions and embeddings endpoints.
    • DeepSeek deepseek-chat gains tools, tool_choice, and stream_options; deepseek-reasoner maps reasoning_effort to thinking mode and returns reasoning_content in streams.
    • Bedrock batch inference now supports embeddings in addition to chat completions.
    • Vertex AI: Portkey metadata maps to Vertex resource labels; enterprise search grounding via enterpriseWebSearch / enterprise_web_search; AWS workloads reach Vertex via AWS–GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys..
    • Azure AI Foundry gains Cohere rerank model support (e.g. cohere.Cohere-rerank-v4.0-pro); gateway strips the cohere. prefix for the provider.

    14 details in Everything

64. updated seen Jul 13, 2026 · issue 014

    Portkey adds vault-backed credentials, weekly rate limits, cross-cloud log delivery, and four new guardrail integrations in March 2026.

    └──▷ USE IT
    Keep LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. provider keys out of Portkey by pointing a virtual key at a secret stored in AWS Secrets Manager — Portkey fetches the value at request time.
    json copy

    {
      "virtual_key": "my-openai-key",
      "secret_mappings": {
        "api_key": "arn:aws:secretsmanager:us-east-1:123456789012:secret:openai-api-key-AbCdEf"
      }
    }

    Gate every request so it must carry a user_id metadata field — the gateway blocks the call before it hits the model if the field is absent.
    json copy

    {
      "guardrails": [
        {
          "type": "required_metadata",
          "required_keys": ["user_id", "session_id"]
        }
      ]
    }

    Apply your existing Zscaler Detection Policy to LLM inputs and outputs by wiring AI Guard into the gateway's request/response hooks.
    json copy

    {
      "guardrails": [
        {
          "type": "zscaler",
          "hooks": ["beforeRequestHook", "afterRequestHook"],
          "policyId": "<your-zscaler-policy-id>",
          "timeout": 10000
        }
      ]
    }

    • Adds Secret References so Portkey fetches credentials at runtime from AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault via secret_mappings, keeping keys out of Portkey directly.
    • Supports weekly budget/usage windows (rpw) and endpoint-scoped rate limits, letting teams cap chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., and admin routes independently.
    • Enables GCS log delivery from AWS-hosted gateways via GCP Workload Identity Federation (GCP_WIF_AUDIENCE / GCP_WIF_SERVICE_ACCOUNT_EMAIL), eliminating long-lived cross-cloud keys.
    • Allows organization admins to include archived workspaces in analytics graphs and filters, with workspace slugs accepted alongside IDs in saved filters.
    • Integrates Zscaler AI Guard as a guardrail partner, applying Zscaler Detection Policies to LLM inputs and outputs via beforeRequestHook / afterRequestHook with a required policyId.

    +8 more show less

    • Integrates Akto Agentic Security as a guardrail partner to scan for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and sensitive data leakage on agentic traffic.
    • Adds customHost on the Bedrock Guardrails plugin so guardrail evaluation can target private or regional Bedrock-compatible endpoints.
    • New required metadata key–value guardrails block requests at the gateway before they reach the model if required metadata fields are missing or invalid.
    • Adds DeepInfra provider support with tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. (tools, tool_choice, parallel_tool_calls), completions, and embeddings endpoints.
    • Expands DeepSeek support: deepseek-chat gains tools, tool_choice, and stream_options; deepseek-reasoner maps reasoning_effort to thinking mode and returns reasoning_content in streams.
    • Extends Bedrock batch inference to support embeddings jobs alongside existing chat completion batch patterns.
    • Adds Vertex AI enterprise web search grounding via enterpriseWebSearch / enterprise_web_search with separate cost attribution, Portkey metadata mapped to Vertex resource labels, and AWS-to-Vertex access via WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys..
    • Supports Azure AI Foundry rerank with Cohere rerank models (e.g. cohere.Cohere-rerank-v4.0-pro), with the gateway stripping the cohere. prefix automatically.

    13 details in Everything

65. Backend 1.20.0 Jul 13, 2026 · release notes

    Portkey Backend 1.20.0 adds Scenario AI provider, Prisma AIRS profile params, SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. domain restriction and atomic OAuth exchange.

    • Adds profile_id, app_name, scan_scope parameters to PANW Prisma AIRS guardrails configuration
    • Makes profile_name optional for PANW Prisma AIRS when using the API key linked profile in Strata Cloud Manager
    • Adds scenario-ai to the Model Catalog for AI-powered image, video, and asset generation
    • Adds disable_auto_provision_user auth setting to restrict SSO login to invited users with verified allowed domains
    • Accepts optional sso_client_id parameter to target specific SSO configurations when linked to multiple setups

    +2 more show less

    • Raises credit_limit in usage limit policies to reactivate already-exhausted keys immediately
    • Enforces single-use atomicity for OAuth authorization-code exchange

    7 details in Everything

66. Enterprise Gateway 2.14.1 Jul 10, 2026 · release notes

    Enterprise Gateway 2.14.1 adds scan_scope and strip_scaffolding for Prisma AIRS agentic support

    • Adds scan_scope parameter to Palo Alto Networks Prisma AIRS guardrail for controlling which messages are scanned (last_message, last_user_message, user_messages, or all_messages)
    • Adds strip_scaffolding parameter to remove known agent-harness wrappers before scanning, reducing false positives for agentic clients

    2 details in Everything

67. Enterprise Gateway 2.14.0 Jul 9, 2026 · release notes

    Portkey Enterprise Gateway 2.14.0 adds HTTP 200 soft deny for guardrails and environment variable limits for config target fan-out.

    • Adds GET /v1/models endpoint to route through the gateway’s provider-backed model listing handler, returning normalized lists with response transforms.
    • Adds soft_deny_200 flag on guardrail checks to return HTTP 200 with error messages formatted as chat completion responses.
    • Adds environment variables MAX_ROOT_CONFIG_TARGETS, MAX_NESTED_CONFIG_TARGETS, and MAX_CONDITIONAL_ROOT_CONFIG_TARGETS to override default config target fan-out limits.

    3 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Config target fan-out is now capped at 100 root-level targets (1,000 for conditional strategy) and 50 nested targets per root target; configs exceeding these limits will be rejected at request time unless overridden by MAX_ROOT_CONFIG_TARGETS, MAX_NESTED_CONFIG_TARGETS, or MAX_CONDITIONAL_ROOT_CONFIG_TARGETS environment variables.
68. Backend 1.19.0 Jul 1, 2026 · release notes

    Portkey Backend 1.19.0 adds Databricks provider support, user-attributed log exports, and stricter workspace security setting overrides.

    • Adds databricks to the Model Catalog with model and pricing configs
    • Rejects workspace security_settings changes for sections not enabled for override
    • Returns next usage-reset timestamp for API key details in Usage Limit Policies
    • Record user-id attribution in log-export routes; User API Keys filter exports by that user
    • Migrates org-level workspace security settings down to workspaces

    5 details in Everything

69. Enterprise Gateway 2.13.0 Jun 30, 2026 · release notes

    Enterprise Gateway 2.13.0 adds server-side MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. execution, Alibaba Cloud AI Guardrails, and tenant-scoped OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. attributes.

    • Adds @portkey-mcp prefix support in the tools array of Responses API and Messages API requests for server-side MCP execution, gated behind the x-portkey-beta: server-side-mcp-2026-06-01 header
    • Adds the alibaba-cloud-ai guardrail plugin for content safety checks, configurable through the standard guardrails workflow
    • Enables MCP gateway request logs to be exported to an OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. collector alongside LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. gateway logs
    • Attaches Portkey tenant attributes (org_id, workspace, environment) to Gen AI and guardrail OTel spans for per-tenant filtering
    • Passes through the thinking.display parameter to the upstream Anthropic API

    5 details in Everything

70. Data Service 1.9.0 Jun 24, 2026 · release notes

    Redis multi-auth and discovery URL support; SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. checks for external calls in Portkey Data Service 1.9.0

    • Adds REDIS_CLUSTER_DISCOVERY_URL environment variable to enable Redis connection through Discovery URL.
    • Adds support for multiple auth modes in Redis improvements for Azure Redis.
    • Adds SSRF checks for external calls.

    3 details in Everything

71. Enterprise Gateway 2.12.0 Jun 23, 2026 · release notes

    Portkey Enterprise Gateway 2.12.0 adds OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. guardrail spans, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface./Agent token metering, Azure MAI image gen, and Vertex MoonshotAI support

    • Guardrail checks now emit dedicated OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. spans under the portkey.guardrail.* namespace for phase, verdict, action, deny, async, and per-check events.
    • Token usage from MCP requests (prompts, completions, tools) is tracked and reported in analytics and logs alongside LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. usage.
    • Token usage from agent endpoints (messages and tasks) is tracked and reported in analytics and logs alongside LLM usage.
    • Image generation and editing are supported for MAI models (e.g., MAI-Image-2.5) on Azure AI Inference.
    • Vertex AI integration now supports the moonshotai model provider.

    +1 more show less

    • The /dataservice/metrics endpoint is gated behind the ENABLE_PROMETHEUS env flag, matching the behavior of the /metrics endpoint.

    6 details in Everything

72. Backend 1.18.0 Jun 23, 2026 · release notes

    Portkey Backend 1.18.0 adds saved-only mode, MiniMax provider, SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. multi-workspace mapping, unified /models endpoint, and JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. agents.invoke scope.

    • Enables block_inline_config data-plane security setting to reject inline provider/endpoint/config selections in favor of admin-curated resources
    • Adds BLOCK_INLINE_CONFIG=true environment variable for self-hosted Gateways to force saved-only mode
    • Adds MiniMax (minimax) model provider to the Model Catalog with model and pricing configs
    • Expands SCIM integration: GET /v1/scim/workspaces now supports pagination (page, page_size, total_count) and groups can map to multiple workspaces
    • Removes include_mapped_groups filter from GET /v1/scim/groups as groups can now map to many workspaces

    +2 more show less

    • Adds fetch_integrated_models config flag and x-portkey-models-endpoint request header to fetch the integrated catalog model list through the Gateway
    • Adds agents.invoke data-plane scope for JWT tokens

    7 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The include_mapped_groups filter is removed from GET /v1/scim/groups endpoint responses
73. Frontend 1.9.3 Jun 22, 2026 · release notes

    Portkey Frontend v1.9.3 adds org-level Data Plane Security controls and Anthropic streaming retry support.

    • Added dashboard controls for Data Plane Security Settings at the organization and workspace levels
    • Added Catch Overloaded Errors on Stream to Anthropic advanced configuration to convert streaming overloaded_error responses to retriable 529 errors

    2 details in Everything

74. Enterprise Gateway 2.11.2 Jun 17, 2026 · release notes

    Adds x-portkey-fetch-integrated-models header to override models API routing, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. metrics-only logging, and S3 log store circuit breaker via env vars.

    • Adds the x-portkey-fetch-integrated-models HTTP header and fetch_integrated_models config field to force GET /v1/models to return gateway-configured models regardless of routing signals
    • Introduces metrics-only logging mode for the MCP gateway, dropping request/response bodies from persisted logs while retaining span info and preserving response bodies on failures

    2 details in Everything

75. Enterprise Gateway 2.11.1 Jun 12, 2026 · release notes

    v2.11.1 updates Bedrock Guardrails action handling, cost attribution, and SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. validation.

    • Guardrails now only block requests when the action field is explicitly blocked, allowing non-blocking actions to pass through.

    1 detail in Everything

76. Enterprise Gateway 2.11.0 Jun 10, 2026 · release notes

    Portkey Enterprise Gateway 2.11 adds MiniMax support, Enforce Inline Config flag, and OAuth forwarding for Claude Code.

    • Adds minimax provider for chat completions with MiniMax’s hosted models
    • Modal integration now supports /v1/embeddings alongside chat completions
    • Anthropic chat completions pass context_management through to the upstream API
    • Enforces inline configs with a new flag that rejects requests passing an inline config JSON object instead of a saved config slug

    4 details in Everything

77. Enterprise Gateway 2.10.0 May 30, 2026 · release notes

    Enterprise Gateway 2.10.0 adds Claude Platform on AWS, Cato guardrail, multi-dimensional pricing, Prometheus labels, and SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. hardening.

    • Adds claude-platform-aws provider with accessKey, assumedRole, and serviceRole auth modes for Anthropic’s full-platform capabilities through AWS
    • Adds Cato Networks guardrail plugin for AI security inspection of LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. prompts and responses via Cato’s analyze API
    • Adds PROMETHEUS_INCLUDE_CONFIG_NAME_LABEL and PROMETHEUS_INCLUDE_API_KEY_NAME_LABEL environment flags to attach config slug and API key name as labels on custom metrics
    • default.allowedRequestTypes guardrail now runs as a beforeRequestHook on proxy requests
    • Adds otel to the plugin manifest allow list to permit OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. export via guardrail config

    +3 more show less

    • guard_chat_completions requests to CrowdStrike AIDR now forward user_id, user_name, llm_provider, and model metadata
    • Vertex AI integration adds support for the zai-org model provider on chat completions
    • Forwarding OpenAI-style service_tier as the upstream Vertex header for tiered inference usage attribution

    8 details in Everything

78. Frontend 1.9.2 May 27, 2026 · release notes

    Portkey Frontend 1.9.2 improves self-hosting configuration via a new environment variable and reorganizes default workspace API key settings to the organisation level.

    • Adds VITE_PRESELECT_LOGIN_PERMISSION environment variable to pre-select the login permission checkbox on private deployments

    1 detail in Everything

79. Backend 1.17.0 May 26, 2026 · release notes

    Adds Cato Networks guardrail, prompt sharing APIs, and org-level API key auto-create settings

    • Introduces Cato Networks guardrail with cato.analyze action supporting monitor, anonymize, or block with userEmail, keyAlias, and timeout options.
    • Adds guardrails.invoke scope for workspace-service and workspace-user API keys.
    • Exposes prompt sharing via new workspace-scoped APIs: POST/PUT/DELETE /v2/prompts/:id/share, /fork, and /unshare.
    • Lists endpoints now support type=shared and show_prompt_share query parameters for filtering shared prompts.
    • Config create/update schema validates default_params and drop_params on targets in addition to override_params.

    +2 more show less

    • Adds org-only setting default_workspace_user_api_key_auto_create to control auto-provisioning of workspace-user API keys.
    • API key rotation enforcement user_api_key_rotation_period now accepts days only (integer 1–365), normalizing old weekly/monthly values.

    7 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Deprecated organisation-level prompt/partial routes removed; use workspace-scoped APIs instead.
80. Enterprise Gateway 2.9.0 May 21, 2026 · release notes

    Enterprise Gateway 2.9.0 adds default/drop params, new providers, Lakera guardrails, and native Azure Responses API

    • Adds default_params and drop_params config fields to Config Targets for conditional injection and removal of request parameters by dot-notation path.
    • Integrates Fal AI provider (fal-ai) for image, video, and audio generation models.
    • Integrates BytePlus provider (byteplus) for chat completions.
    • Integrates Scenario AI provider (scenario-ai) with proxy passthrough support for Basic auth using apiKey and optional scenarioApiSecret.
    • Adds Lakera Guard as a guardrails partner via /v2/guard API, supporting beforeRequestHook/afterRequestHook, PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. redaction, and optional project_id/apiBase.

    +4 more show less

    • Enables per-region Vertex AI endpoints for batches, Gemini EmbeddingEmbeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. 2 embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., Anthropic Claude on Vertex, and Gemini models.
    • Adds support for Anthropic Claude models via the unified /v1/batches API on Vertex AI.
    • Unlocks Azure AI Foundry native Responses API (/v1/responses) behind a flag for agent references and transformations.
    • Supports OpenAI-style service_tier parameter in Amazon Bedrock provider requests with correct usage attribution.

    9 details in Everything

81. Frontend 1.9.1 May 21, 2026 · release notes

    Frontend 1.9.1 adds Azure Workload Identity configuration via dashboard and improves API key rotation UX.

    • Adds azureWorkloadClientId and azureWorkloadTenantId config fields in the dashboard to enable Azure Workload Identity for OpenAI and AI Foundry integrations
    • Improves input field and validation error messaging for Org service API key rotation

    2 details in Everything

82. Frontend 1.9.0 May 20, 2026 · release notes

    Frontend 1.9.0 adds org workspace publishing, CSV variable import/export, Scenario AI provider support, and a Shared with me tab.

    • Publish to org or selected workspaces for prompt and partial sharing to control visibility
    • Shared with me tab displays org-wide shared prompts and partials
    • Export prompt variables as CSV (name, type, value) and import from CSV in bulk
    • New provider Scenario AI added

    4 details in Everything

83. Data Service 1.8.0 May 20, 2026 · release notes

    Data Service 1.8.0 adds provider batch guardrails and opt-in Bull Board dashboard.

    • Adds input_guardrails and output_guardrails config keys inside portkey_options to enable redaction/filtering on provider batches.
    • Adds BULL_BOARD_ENABLED=true environment variable to enable the Bull Board queue dashboard (disabled by default).

    2 details in Everything

84. Backend 1.16.2 May 19, 2026 · release notes

    Portkey Backend 1.16.2 adds Lakera Guard support, workspace API key defaults, batch API key lookup, and SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. hardening.

    • Adds Lakera Guard integration via /v2/guard endpoint for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools., policy violation, and PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. detection with projectID, endInclusive, and timeout config options
    • Adds fal.ai provider to the Model Catalog under the fal-ai identifier for OpenAI-compatible inference across LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code., image, video, and audio models
    • Allows setting defaults.user_api_key_config via PUT /v2/workspaces/:id to define a default config slug for auto-created or fallback workspace-user API keys, with validation preventing deletion of the active default
    • Extends isValidCustomHost validation across organisation settings, deployments, providers, virtual keys, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. integrations, integration-model custom_host, and SAMLSAMLA standard for passing signed identity assertions between a sign-in provider and an application. It lets a user sign in through their organization's identity provider. ACS URLs to block private/reserved IPs, metadata endpoints, non-HTTP(S) schemes, URL fragments, and other unsafe targets
    • Enables hardening of OIDCOIDCOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user. PKCEPKCEProof Key for Code Exchange, an OAuth safeguard that binds an authorization code to the client that requested it. An intercepted code cannot be redeemed without the client's matching verifier. flows for SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. authentication and ensures JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. user attribution correctly resolves user email and enforces workspace membership

    +2 more show less

    • Adds ids query parameter support to GET /v2/api-keys for batch lookup via comma-separated UUIDsUUIDsA universally unique identifier, usually a 128-bit value used to identify an object across systems. It is designed to make independently generated identifiers unlikely to collide.
    • Validates byteplus_region in the integrations API for the BytePlus provider integration

    7 details in Everything

85. Enterprise Gateway 2.8.0 May 14, 2026 · release notes

    Portkey Enterprise Gateway 2.8.0 adds Tavily guardrails, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. CORSCORSA browser mechanism that lets a server declare which foreign origins may read its HTTP responses, enforced by the browser to block unauthorized cross-site requests from JavaScript. controls, Azure Workload Identity, and Vertex AI TTS support.

    • Adds Tavily Online Guardrail Plugin for web-grounded checks in guardrail pipelines
    • Enables Vertex AI Gemini TTS via unified speech APIs (createSpeech / generateSpeech) through gateway routes
    • Updates Bedrock Mantle prompt cache pricing to include 1-hour cached context cost path
    • Extends Workload Identity support for Azure OpenAI, Azure AI Foundry routes and adds it as an authentication option for Azure Cache for Redis
    • Introduces a new usage limit policy type for finer-grained consumption control

    +4 more show less

    • Adds environment flags to tune CORS response headers for MCP endpoints in self-hosted deployments
    • Extends Prompt Security checks with additional coverage and provider-specific configuration hooks
    • Refreshes Beta header forwarding and filtering across providers
    • Supports provider slug references for circuit breaker target resolution

    9 details in Everything

86. Backend 1.16.1 May 14, 2026 · release notes

    Portkey 1.16.1 adds SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly.-hardened Agent Gateway fetching, batches output file ID support, and service-to-service config auth

    • POST /v2/agent-integrations/fetch-agent-card now validates upstream endpoints against SSRF vectors (private IPs, cloud metadata, non-HTTP(S), fragments) and enforces valid JSON {url, name} responses under 1 MB
    • Agent card discovery is restricted to /.well-known/agent-card.json and /.well-known/agent.json, removing the previous fallback to the raw base endpoint
    • PUT /v2/batches/:id now accepts internal_output_file_id for the guardrails-on-batches pipeline, supported by a new internal column on batches records
    • GET /v2/configs now supports data-service token authentication for service-to-service configuration reads

    4 details in Everything

87. Backend 1.16.0 May 13, 2026 · release notes

    Portkey Backend 1.16.0 adds Tavily guardrail, Azure Workload Identity, requests usage-limit policy type, and configurable API key rotation.

    • Adds Tavily Online Search guardrail with configurable search parameters to inject web results into prompts
    • Adds workload auth mode for Azure OpenAI and Azure AI Foundry integrations requiring azure_workload_client_id and azure_workload_tenant_id
    • Adds entraFederated auth mode for Azure AI Foundry integrations
    • Adds requests as a new usage-limits policy type alongside cost and tokens
    • Allows PUT /v2/policies/usage-limits/:id and PUT /v2/policies/rate-limits/:id to accept conditions in the request body

    +3 more show less

    • Adds rotation_period_days (1–365) to API key rotation policies alongside existing rotation_period
    • Honours NO_PROXY environment variable for outbound LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code., OIDCOIDCOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user., and Mailgun calls
    • Adds REQUEST_TIMEOUT config for outbound fetch timeout in milliseconds

    8 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !GET /v1/auth/self now requires a session/JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. and rejects requests authenticated only with an API key
    • !Legacy endpoints /users/firebase-custom-token and /utils/oidc-client-id have been removed
88. Frontend 1.8.2 May 12, 2026 · release notes

    Portkey Frontend 1.8.2 adds auto API key creation, budget reset, and policy condition updates.

    • Adds the ability to automatically create an API key for an organisation and workspace
    • Adds the ability to reset usage for budget allocation
    • Enables conditions update for policy updates

    3 details in Everything

89. Data Service 1.7.1 May 12, 2026 · release notes

    Portkey Data Service adds Azure Blob Storage backend and FIPSFIPSA set of US federal standards published by NIST that define approved cryptographic algorithms and modules; cyber tools use it to meet compliance requirements in government and regulated environments.-compliant Dockerfile variant in v1.7.1

    • Adds Azure Blob Storage as a log-export storage backend alongside existing S3 options.
    • Adds FIPS-compliant Dockerfile variant for data-service with FIPS mode enabled and SHA-256SHA-256A cryptographic hash function that turns data into a fixed 256-bit fingerprint. It is used to check integrity; hashing does not encrypt the original data. signing.

    2 details in Everything

90. Backend 1.15.0 May 4, 2026 · release notes

    Portkey Backend v1.15 adds auto-generated API keys, custom key rotation, passthrough configs, and granular agent RBACRBACRole-based access control, where permissions are assigned to roles such as reader or administrator and users receive those roles. It makes access easier to manage than individual permission lists..

    • Adds user_api_key_auto_create organization setting to automatically provision default workspace-user API keys on member join or invite acceptance
    • Adds user_api_key_rotation_period organization setting to enforce automatic API key rotation for workspace-user keys
    • Adds rotation_period_days column to rotation_policies and exposes periodic_reset_days, next_usage_reset_at, and last_reset_at fields on usage limit policies
    • Adds API_KEY_PREFIX environment variable to prefix all newly-generated Portkey API keys in self-hosted deployments
    • Adds pricing_adjustments.multiplier config to integrations for per-token markup/discount multipliers

    +13 more show less

    • Allows setting models[].configurations.custom_host per integration model via PUT /integrations/:id/integration-models
    • Supports value_format: json in secret_mappings to parse stringified JSON values
    • Adds support for secret_mappings on MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. integrations with resolution at request time
    • Adds azure_auth_mode: entraFederated mode for Azure OpenAI integrations, requiring azure_entra_tenant_id and azure_entra_client_id without a client secret
    • Seeds new bedrock-mantle provider for AWS's OpenAI-compatible Bedrock inference engine
    • Adds allow_subdomains flag to SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. settings to resolve users from subdomains of an allowed root domain
    • Extends POST /v1/scim/workspaces to accept scim_group_name in addition to scim_group_id
    • Adds organisationAdminsViewLogs and organisationAdminsViewLogMetadata organization settings to control org admin access to request logs and metadata
    • Allows configs to set passthrough: true to mark requests as passthrough mode without provider transformation
    • Adds matchType parameter (exact, contains, containsAll, regex) for guardrails metadata key-value comparison
    • Makes /models route accessible to all workspace roles and renames output field total to total_count
    • Adds granular RBAC scopes: organisation_agent_integrations.*, workspace_agent_integrations.*, agent_servers.*, agent_server_skills.*, agent_integration_skills.*, agent_server_user_access.*, and agents.invoke
    • Adds GET /v1/agent-servers/:id/agent-card endpoint to fetch the upstream agent card

    18 details in Everything

91. Enterprise Gateway 2.7.1 Apr 30, 2026 · release notes

    Portkey Enterprise Gateway 2.7.1 adds anthropicInspectStreamForOverloadedError to handle Anthropic streaming errors.

    • Adds anthropicInspectStreamForOverloadedError configuration key to inspect the first chunk of an Anthropic streaming response and return HTTP 529 on overloaded_error, enabling proper retry/fallback logic

    1 detail in Everything

92. Enterprise Gateway 2.7.0 Apr 30, 2026 · release notes

    Portkey Gateway v2.7 adds Vertex AI Rerank, Bedrock Mantle provider, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. secret support, and Sigstore image signing

    • Adds bedrock-mantle provider for AWS OpenAI-compatible Bedrock inference with apiKey, assumedRole, and serviceRole auth modes via /v1/chat/completions, /v1/responses, and /v1/messages
    • Supports Vertex AI Rerank on the unified /v1/rerank endpoint using semantic-ranker-default, semantic-ranker-fast, or semantic-ranker-512 models
    • Adds matchType parameter to Required Metadata Key-Value Pairs guardrail supporting exact, contains, containsAll, and regex comparison behaviors
    • Enables MCP integrations to reference centrally-managed secrets via secret_mappings in the integration config, resolved at connection time
    • Adds value_format: json support for Secret Mappings to parse stringified JSON into target fields for structured credentials

    +3 more show less

    • Allows integration-level cost multipliers via pricing_adjustments.multiplier config key for per-integration markup or discount pricing
    • Outputs signed Docker images using Sigstore Cosign, verifiable against Portkey’s public key before deployment
    • Extends SEMANTIC_CACHE_EMBEDDING_PROVIDER to accept google and vertex-ai values alongside openai for semantic cache population

    8 details in Everything

93. Frontend 1.7.6 Apr 24, 2026 · release notes

    Frontend 1.7.6 adds guardrail pagination, Azure OpenAI/Foundry secret reference integration, and performance fixes

    • Guardrail List page now supports pagination to handle large numbers of guardrails.
    • Secret Reference Integration support for Azure OpenAI and Azure Foundry setup inside LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Integrations.

    2 details in Everything

94. Frontend 1.7.5 Apr 22, 2026 · release notes

    Frontend 1.7.5 adds Org/Workspace log management permissions and granular guardrail view/write toggles

    • Adds Org-Admin permissions for Logs Management at both Org level and Workspace level
    • Adds per-role view and write toggles for Guardrail Management security settings with workspace-level override

    2 details in Everything

95. Enterprise Gateway 2.6.2 Apr 17, 2026 · release notes

    Portkey Gateway 2.6.2 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. header renaming, Azure Entra federated auth, and Gemini EmbeddingEmbeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. 2 support

    • forward_headers entries now accept { from, to } objects to rename client headers before they reach the upstream.
    • Adds entraFederated auth mode for Azure OpenAI/AI Foundry to exchange AWS web identity tokens for Entra IDEntra IDMicrosoft's cloud-based identity and access management service, formerly Azure Active Directory, used by cyber tools to authenticate users and enforce access policies in Microsoft 365 and Azure environments. access tokensaccess tokensA credential that authorizes a client to use a service or API. It usually has limited permissions and a lifetime; anyone holding it may be able to use those permissions..
    • gemini-embedding-2-preview is now available on the unified /v1/embeddings route for Vertex AI.

    3 details in Everything

96. Frontend 1.7.4 Apr 16, 2026 · release notes

    Portkey Frontend 1.7.4 adds Logs Path variable in UI, Metadata to Vertex Labels mapping, custom model host provisioning, and Entra-Auth federated support for Azure.

    • Logs Path variable is now available in the UI
    • Maps Portkey Metadata to Vertex Labels for better observability
    • Enables federated auth support for Entra-Auth for Azure Integrations

    3 details in Everything

97. Frontend 1.7.3 Apr 14, 2026 · release notes

    Adds UI support for Integrations-specific Security Settings and Policy type exclusion in Frontend 1.7.3

    • Updates UI to support excluding Policy types

    1 detail in Everything

98. Enterprise Gateway 2.6.1 Apr 14, 2026 · release notes

    Adds Anthropic Batch API support, sensitive header masking, Bedrock reasoning effort, and Agent Gateway preview

    • Adds /v1/batches endpoint for running asynchronous Claude jobs via the Anthropic Batch API
    • Adds x-portkey-sensitive-headers request header to mask matching values in logs and traces
    • Adds support for reasoning_effort parameter (minimal/low/medium/high/none) for Bedrock Claude reasoning models, mapped to thinking.budget_tokens on earlier models
    • Adds experimental /v1/agent/:agentServerId/* proxy route for Agent Gateway preview with A2AA2AAgent2Agent, an open protocol started by Google for AI agents from different vendors to discover each other, delegate tasks and exchange results. It complements MCP, which connects an agent to tools.-registry support

    4 details in Everything

99. Backend 1.14.0 Apr 12, 2026 · release notes

    Portkey Backend 1.14.0 introduces API key rotation, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. connection APIs, Lasso classifier, and HADB MySQL support

    • API Key Rotation is generally available; use *.rotate scope for organization-service, workspace-service, and workspace-user API key types to enable rotation
    • Added organization security settings managersWriteWsIntegrations, managersWriteWsMcpIntegrations, and managersWriteMcpServers with default enabled and workspace-level overrides to control manager write access
    • New GET /v1/mcp-servers/:mcpServerId/connections API for paginated list of user connections to an MCP server
    • New DELETE /v1/mcp-servers/:mcpServerId/connections API to revoke a specific user's connection to an MCP server
    • Added configurable MySQL reader with HADR mode using new env vars DB_READER_HOST, DB_READER_USER, DB_READER_PASS, DB_READER_PORT and MYSQL_WRITER_HADR_MODE=ON for writer failover

    +1 more show less

    • Added new lasso.classify check to Guardrails Lasso integration with optional apiEndpoint field for self-hosted deployments

    6 details in Everything

100. Enterprise Gateway 2.6.0 Apr 8, 2026 · release notes

     Portkey Enterprise Gateway v2.6.0 adds Vertex AI label mapping, custom model host routing, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. timeout/env vars, and data URL support.

     • Adds automatic mapping of Portkey metadata to Vertex AI resource labels across chat/completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., batches, and fine-tuning with configurable METADATA_MAP_KEY_PREFIX environment variable
     • Enables custom_host routing in model config for custom models, allowing different upstreams per model under the same integration while respecting header-level priority
     • Adds support for base64 data URLs (format: data:<media_type>;base64,<data>) in file.file_data via /chat/completions to forward as Anthropic document content
     • Enables passthrough of top-level cache_control parameter to Anthropic in /chat/completions and adds automatic additionalProperties: false for strict structured outputsstructured outputsForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed.
     • Adds configurable outbound timeout for MCP calls via MCP_REQUEST_TIMEOUT_MS environment variable and updates MCP_POOL_* env var precedence with new log attributes gateway_version and mcp.auth.upstream_token_refreshed

     5 details in Everything

101. Enterprise Gateway 2.5.1 Apr 1, 2026 · release notes

     Output guardrails now apply to streaming responses for /chat/completions, /completions, and /messages endpoints.

     • Adds output_guardrails and after_request_hooks support for streaming responses via SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. chunks at stream end

     1 detail in Everything

102. Frontend 1.7.0 Apr 1, 2026 · release notes

     Frontend 1.7.0 adds Gateway Registration, API Key Rotation UIs, Anthropic Cache TTLTTLTime to live, a limit on how long data remains valid or how far a packet can travel. Caches use a time limit; IP packets use a hop limit. support, Akto AI plugin, and enhanced Trace View filtering.

     • Gateway Registration UI added for configuring gateway connections
     • API Key Rotation UI added for managing key lifecycle
     • New Last Reset Date field displayed in Integrations, Providers, API Keys, and Workspace Budgets views
     • Trace View now supports displaying spans without active filters via a new toggle UI element
     • Added configuration option for Anthropic Cache TTL from Playground/Prompts settings

     +2 more show less

     • Organisation-level maximum session time setting now supported for user management
     • Akto AI added as a supported plugin within the system

     7 details in Everything

103. Backend 1.13.0 Apr 1, 2026 · release notes

     Portkey Backend v1.13.0 adds API key rotation, organization session management, Required Metadata guardrails, and Agent Gateway CRUDCRUDThe four basic database operations — Create, Read, Update, Delete — used as shorthand for any system that manages persistent data records, giving cyber tools a standard model for interacting with stored information. APIs.

     • Adds automatic and manual API key rotation with rotation_policy config (rotation_period, next_rotation_at, key_transition_period_ms) and new POST /api-keys/:apiKeyId/rotate endpoint
     • Adds session_settings organization-level auth configuration with max_session_ttl to reject expired user sessions
     • Adds Required Metadata Key-Value Pairs guardrail (default.requiredMetadataKeyPairs) with metadataPairs and operator parameters in beforeRequestHook
     • Adds rpw (requests per week) rate limit unit and endpoint_type condition key for granular rate limiting
     • Adds Agent Gateway with full CRUD APIs, agent skills management, workspace-level access control, and RBACRBACRole-based access control, where permissions are assigned to roles such as reader or administrator and users receive those roles. It makes access easier to manage than individual permission lists. support

     5 details in Everything

104. Enterprise Gateway 2.5.0 Apr 1, 2026 · release notes

     Gateway-Local JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. auth via JWT_ENABLED, HEADERS_TO_METADATA injection, weekly rate limits, metadata guardrail, Lasso v3 API, HEIC/HEIF images

     • Adds JWT_ENABLED=ON environment variable to enable gateway-local JWT validation and support resolving Organisation ID from portkey_oid or organisation_id claims.
     • Adds HEADERS_TO_METADATA environment variable to inject request headers into metadata for observability enrichment without client-side x-portkey-metadata changes.
     • Adds weekly window rate limiting via rpw and endpoint type targeting via endpoint_type condition key in rate limit policies.
     • Adds native metadataPairs guardrail check with all, any, and none operators to validate metadata before request processing.
     • Upgrades Lasso Security integration to v3 API Classify, enabling structured findings-based detection, session/user tracking via sessionId/userId, and configurable apiEndpoint.

     +1 more show less

     • Adds support for image/heic and image/heif MIME types in multimodalmultimodalAble to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. requests and the Inline Image URLs guardrail.

     6 details in Everything

105. April 2026 Apr 2026 · release notes

     Portkey introduces Agent Gateway CRUDCRUDThe four basic database operations — Create, Read, Update, Delete — used as shorthand for any system that manages persistent data records, giving cyber tools a standard model for interacting with stored information., Lasso v3 structured findings, streaming output guardrails, and several gateway/model provider updates for April 2026.

     • Adds Agent Gateway with CRUD APIs for agent integrations, agent servers, skills, and RBACRBACRole-based access control, where permissions are assigned to roles such as reader or administrator and users receive those roles. It makes access easier to manage than individual permission lists. access control
     • Enables API key rotation on schedule or demand, including an explicit manual rotate endpoint and usage reset alongside rotation
     • Introduces Organization session policy with the max_session_ttl setting to enforce idle session expiration
     • Enables local JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. validation in the gateway by setting the environment variable JWT_ENABLED=ON
     • Adds HEADERS_TO_METADATA config key to copy inbound headers into request metadata for tracing and analytics

     +17 more show less

     • Supports weekly rate limit windows via the rpw policy type and endpoint-scoped limits by type (e.g., chat vs. embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.)
     • Adds gateway routing configs: passthrough targets, sticky session strategy (strategy.sticky), forward header mapping (forward_headers), and subdomain allowance (allow_subdomain)
     • Enables streaming output guardrails that return a final chunk with the check outcome on the same stream
     • Adds capability to require specific metadata key-value pairs on every LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. request, rejecting mismatches before forwarding
     • Upgrades Lasso security checks to v3 API, returning structured findings (block, auto-mask, warn) and supporting sessionId/userId ties across endpoints
     • Adds Model Rules to tie model access to request metadata and a request-params check to govern tools and parameters before traffic leaves Portkey
     • Separates guardrail definition permissions from execution permissions and clarifies org-level full request log access controls
     • Adds Partner guardrails for building guardrails as partner integrations with built-in step-by-step guidance
     • Guides semantic cache setup directly in the Portkey app UI with in-product policy checks
     • Adds support for Anthropic Files API and batch workloads, including service_tier mapping to speed/priority tiers
     • Enables calling the rerank endpoint through the same Portkey path as chat and embeddings
     • Adds routing for Gemini embedding modelsembedding modelsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. (including 2-preview) using standard OpenAI-compatible shapes
     • Adds support for image/heic and image/heif in multimodalmultimodalAble to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. requests, with Inline Image URL guardrail support
     • Preserves the response_format field for Mistral structured JSON outputs through the gateway
     • Adds dedicated connection flows for Claude Cowork, Hermes Agent, Pi, and Slack MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. integrations
     • Surfaces Skills Registry across Portkey in place of the previous Agent Skills naming
     • Updates self-hosted deployment images and provisioning for GCP, AWS, Azure, and Azure Container Apps

     22 details in Everything

106. Enterprise Gateway 2.4.4 Mar 25, 2026 · release notes

     Portkey Enterprise Gateway 2.4.4 adds Anthropic Files API support via /v1/files, cache TTLTTLTime to live, a limit on how long data remains valid or how far a packet can travel. Caches use a time limit; IP packets use a hop limit., and OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. semconv alignment.

     • Adds EXPERIMENTAL_GEN_AI_OTEL_RESOURCE_ATTRIBUTES environment variable for custom OTLPOTLPOpenTelemetry Protocol, a standard way to send traces, metrics and logs between applications, collectors and monitoring systems. It lets telemetry move across compatible tools. resource attributes
     • Supports the /v1/files endpoint for Anthropic Files API integration
     • Adds optional ttl field to cache_control (values 5m or 1h) for Prompt Caching

     3 details in Everything

107. Enterprise Gateway 2.4.3 Mar 17, 2026 · release notes

     Portkey Enterprise Gateway 2.4.3 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. event observability and Prometheus env var support

     • Adds PROMETHEUS_INCLUDE_MODEL_LABEL, PROMETHEUS_INCLUDE_METADATA_LABELS, PROMETHEUS_HISTOGRAM_BUCKETS, PROMETHEUS_HTTP_DURATION_BUCKETS, and PROMETHEUS_GC_BUCKETS environment variable support for configuring Prometheus metrics.
     • Enhances MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. logging to capture all MCP events, including caller identity, authorization scope, decision, and gateway trace_id.

     2 details in Everything

108. Enterprise Gateway 2.4.2 Mar 16, 2026 · release notes

     Enterprise Gateway v2.4.2 adds native Bedrock structured outputstructured outputForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. enforcement via Converse API's json_schema.

     • Bedrock structured outputsstructured outputsForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. now use outputConfig with json_schema text format via the Converse API for Claude models on Bedrock

     1 detail in Everything

109. Backend 1.12.1 Mar 16, 2026 · release notes

     Portkey Backend v1.12.1 adds Akto Agentic Security guardrail, inline image URL processing, and Bedrock Custom Host support.

     • Adds Akto Agentic Security guardrail integration for advanced threat detection and security scanning of LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. inputs and outputs
     • Adds Inline Image URLs guardrail that fetches external image URLs and converts them to Base64 inline data, with configurable providers, maxSizeBytes (default: 20MB), timeoutMs (default: 30000), and failOnError (default: false)
     • Adds optional Custom Host field to the AWS Bedrock guardrail integration schema

     3 details in Everything

110. Enterprise Gateway 2.4.0 Mar 11, 2026 · release notes

     Portkey Enterprise Gateway 2.4.0 adds DeepInfra tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. and Vertex AI metadata forwarding

     • Adds tool_choice and parallel_tool_calls parameters to the DeepInfra integration for function callingfunction callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. workflows
     • Adds support for the completions and embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. endpoints in the DeepInfra provider configuration
     • Forwards Portkey request metadata as Vertex AI resource labels to enable enterprise governance

     3 details in Everything

111. Backend 1.12.0 Mar 9, 2026 · release notes

     Portkey Backend 1.12.0 adds Secret Manager integrations, Zscaler guardrail, and GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. log storage.

     • Adds secret_mappings config key and full CRUDCRUDThe four basic database operations — Create, Read, Update, Delete — used as shorthand for any system that manages persistent data records, giving cyber tools a standard model for interacting with stored information. API for managing external secret references to AWS Secrets Manager, Azure Key Vault, and HashiCorp Vault
     • Adds new Zscaler AI Guard integration via beforeRequestHook and afterRequestHook with policyId and timeout parameters
     • Adds GCP Workload Identity Federation support for cross-cloud log storage using gcs_assume type and env vars GCP_WIF_AUDIENCE / GCP_WIF_SERVICE_ACCOUNT_EMAIL

     3 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !Requires Helm chart upgrade to app-1.6.0 or higher to function correctly
112. Enterprise Gateway 2.3.0 Mar 6, 2026 · release notes

     Portkey Enterprise Gateway 2.3.0 adds Akto guardrails, DeepSeek tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions./reasoning, Azure rerank, Vertex web search, AWS-to-GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys., and Bedrock batch embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG..

     • beforeRequestHook and afterRequestHook now support the Akto Agentic Security guardrail with a configurable timeout (default 5000ms) for prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. and data leakage protection
     • deepseek-chat provider gains support for tools, tool_choice, and stream_options parameters, enabling tool calling capabilities
     • deepseek-reasoner provider adds reasoning mode via the reasoning_effort parameter and returns reasoning_content in streaming responses
     • Azure AI Foundry now supports rerankingrerankingA model that re-scores a first batch of search results against the query and reorders them, more accurately than the fast first-stage search. via Cohere models (e.g., cohere.Cohere-rerank-v4.0-pro) with automatic stripping of the cohere. prefix
     • Vertex AI provider accepts enterpriseWebSearch or enterprise_web_search in the tools array to enable enterprise web search grounding with distinct cost attribution

     5 details in Everything

113. Frontend 1.6.8 Mar 6, 2026 · release notes

     Portkey Frontend 1.6.8 introduces Zscaler integration and Policy v2.0.0 support.

     • Introduced Zscaler integration
     • Upgraded Policy to version 2.0.0

     2 details in Everything

114. Frontend 1.6.7 Mar 2, 2026 · release notes

     Adds conditional redirection and new security reference feature to Playground UI

     • Added conditional redirection support
     • Added summary tables for improved data visibility
     • Added new security reference feature

     3 details in Everything

115. March 2026 Mar 2026 · release notes

     Portkey March 2026 adds secret references, zscaler/akto guardrails, GCP WIFWIFWorkload Identity Federation is a Google Cloud mechanism that lets external workloads authenticate to Google APIs using short-lived tokens from a trusted identity provider, eliminating the need to manage long-lived service account keys. logs, and new model support

     • Adds secret_mappings configuration to reference credentials from AWS Secrets Manager, Azure Key Vault, or HashiCorp Vault instead of storing keys directly.
     • Adds rpw weekly budget windows for rate limiting, allowing spend caps aligned with weekly review cycles.
     • Adds endpoint-scoped rate limits by endpoint type (e.g., chat completions, embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.) to apply different limits to different API surfaces.
     • Adds support for Zscaler AI Guard as a guardrails partner via beforeRequestHook and afterRequestHook with required policyId.
     • Adds support for Akto Agentic Security as a guardrails partner for scanning LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. inputs/outputs, with configurable timeout (default 5000 ms).

     +10 more show less

     • Adds customHost field to Bedrock guardrail plugin configuration to direct checks to private or regional endpoints.
     • Adds required metadata key–value enforcement in guardrails to block requests missing specified fields before model reach.
     • Adds GCS log storage capability from AWS using GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL for cross-cloud authentication without static keys.
     • Adds workspace slug support in analytics filters alongside IDs for archived workspaces.
     • Adds DeepInfra provider support with tools, tool_choice, and parallel_tool_calls for tool callingtool callingA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions., plus completions and embeddings endpoints.
     • Adds DeepSeek (deepseek-chat, deepseek-reasoner) provider support including stream_options handling and reasoning content mapping.
     • Adds Vertex AI enterprise web search grounding via enterpriseWebSearch/enterprise_web_search with separate cost attribution.
     • Adds AWS–GCP WIF support for Vertex AI access using GCP_WIF_AUDIENCE and GCP_WIF_SERVICE_ACCOUNT_EMAIL.
     • Adds Azure AI Foundry rerank support for Cohere models, stripping the cohere. provider prefix.
     • Adds Bedrock batch inference support for embeddings alongside chat completions.

     15 details in Everything

116. Backend 1.11.2 Feb 27, 2026 · release notes

     Portkey Backend 1.11.2 adds grouped analytics endpoints, archived workspace filtering, and API enhancements for integrations, prompts, and gateway config.

     • Added GET /v1/analytics/groups/:groupBy and GET /v1/logs/groups/:groupBy generic grouped analytics endpoints with configurable columns
     • Added status query parameter to the workspace list API (GET /workspaces) to filter by active or archived status
     • Added create_default_provider and default_provider_slug fields to the integration creation API for workspace integrations
     • Added patch flag to PUT /prompts/{promptId} to enable partial version field updates
     • Added azure_entra_scope option to the gateway config schema for specifying custom Azure Entra IDEntra IDMicrosoft's cloud-based identity and access management service, formerly Azure Active Directory, used by cyber tools to authenticate users and enforce access policies in Microsoft 365 and Azure environments. authentication scopes at the config level

     5 details in Everything

117. Enterprise Gateway 2.2.5 Feb 27, 2026 · release notes

     Enterprise Gateway 2.2.5 adds secondary Redis for workers and configurable Prometheus metrics.

     • Adds REDIS_WORKERS_URL, REDIS_WORKERS_HOST, REDIS_WORKERS_PORT, REDIS_WORKERS_USERNAME, REDIS_WORKERS_PASSWORD, REDIS_WORKERS_TLS_ENABLED, and REDIS_WORKERS_AUTH_MODE environment variables to route BullMQ workers to a dedicated Redis instance.
     • Adds PROMETHEUS_INCLUDE_MODEL_LABEL, PROMETHEUS_INCLUDE_METADATA_LABELS, and PROMETHEUS_LABELS_METADATA_ALLOWED_KEYS environment variables to configure Prometheus metric cardinality.
     • Adds native Responses API support for the Perplexity provider.
     • Adds embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. endpoint support for the OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code. provider.

     4 details in Everything

118. Backend 1.11.1 Feb 24, 2026 · release notes

     Portkey Backend 1.11.1 introduces configuration capabilities for airgapped environments, specifically allowing custom OpenID ConnectOpenID ConnectOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user. scopes via a new environment variable.

     • Adds OIDC_CUSTOM_SCOPES environment variable to configure additional custom OIDCOIDCOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user. scopes (comma-separated) for airgapped deployments

     1 detail in Everything

119. Frontend 1.6.6 Feb 23, 2026 · release notes

     Portkey Frontend adds API key filtering by type and column sorting

     • Adds the ability to filter API keys by API Key Type (User, Service) in the frontend interface
     • Adds the ability to sort API keys by column name in the frontend table view

     2 details in Everything

120. Backend 1.11.0 Feb 20, 2026 · release notes

     Portkey Backend 1.11.0 adds data_visibility workspace settings for scoped observability data access.

     • Adds membersViewAllData setting to restrict workspace members' view of logs and analytics to their own keys
     • Adds managersViewAllData setting to restrict managers' view of logs and analytics to the keys they manage
     • Supports data_visibility override category for workspace-level configuration

     3 details in Everything

121. Enterprise Gateway 2.2.4 Feb 20, 2026 · release notes

     Portkey Enterprise Gateway 2.2.4 adds Anthropic Bedrock citations, Zscaler AI Guard integration, SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. streaming for TTS, and ZhipuAI image generation support.

     • Adds stream_format: "sse" to OpenAI and Azure OpenAI text-to-speech requests for Server-Sent Events audio streaming with usage logging
     • Adds Zscaler AI Guard as a new guardrails partner for DLPDLPData Loss Prevention: a category of controls that inspect and block sensitive data from leaving an organisation via email, web, or removable media, giving cyber tools a policy enforcement layer over data in motion and at rest. and prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. protection on inbound prompts and outbound responses
     • Adds Anthropic’s citations feature support on Bedrock for chat completions API
     • Adds ZhipuAI image generation support through the CogView model family (e.g., cogview-4-250304) with parameters prompt, model, n, size, response_format, and user
     • Adds x-portkey-vertex-auth-type provider option for Vertex AI to configure authentication type (e.g., workload identityworkload identityAn identity assigned to software, such as a service or container, rather than a person. It lets the workload authenticate and receive permissions without sharing a user's credentials.)

     +1 more show less

     • Automatically records user_id in analytics from the API key’s associated user

     6 details in Everything

122. Backend 1.10.0 Feb 18, 2026 · release notes

     Backend 1.10.0 adds prompt access control settings, custom budget reset intervals via new API fields, and SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. group management for AirGapped environments.

     • membersViewPrompts and membersWritePrompts config keys to control workspace member prompt permissions
     • managersViewPrompts and managersWritePrompts config keys to control workspace manager prompt permissions
     • Custom periodic reset intervals for workspace budget limits via the periodic_reset_days field (1–365 days)
     • The next_usage_reset_at field (ISOISOA widely adopted international standards body that publishes technical specifications; cyber tools cite ISO standards to signal compliance with recognized benchmarks for quality, security, or interoperability. 8601) to specify the first custom reset date for budgets
     • SCIM User-Based Group Management in AirGapped mode using the SCIM /Users endpoint when SCIM_MEMBERSHIP_USER_MODE=ON is set

     +2 more show less

     • groups attribute included in user SCIM responses to expose current group memberships
     • New authentication type Workload Identity Federation for Vertex AI integrations

     7 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !periodic_reset_days is mutually exclusive with the existing periodic_reset (weekly / monthly) option
123. Frontend 1.6.3 Feb 18, 2026 · release notes

     Frontend 1.6.3 adds API key filtering, prompts security settings, and workspace budget resets

     • Adds support for User and Service key filter in API Key List page
     • Adds Security setting for Prompts View and Management
     • Adds support for Workspace Budget periodic reset by date

     3 details in Everything

124. Backend 1.9.0 Feb 16, 2026 · release notes

     Portkey Backend 1.9.0 adds per-role analytics access control, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. metadata APIs, and self-hosted feature flags.

     • Added managersViewAnalytics config setting to control workspace manager analytics visibility
     • Added membersViewAnalytics config setting to control workspace member analytics visibility
     • Added GET /mcp-servers/:id/metadata endpoint for fetching detailed MCP server metadata

     3 details in Everything

125. Enterprise Gateway 2.2.3 Feb 13, 2026 · release notes

     Enterprise Gateway 2.2.3 adds Databricks provider, Together AI reasoning models, and native Responses API for Groq/OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code./xAI

     • Adds Databricks as a new provider for chat completions, completions, and embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG..
     • Enables Together AI reasoning/thinking model support via the reasoning_effort parameter and content_blocks in responses when strictOpenAiCompliance is set to false.
     • Routes Groq, OpenRouter, and xAI Responses API requests to their native /responses endpoints.

     3 details in Everything

126. Enterprise Gateway 2.2.2 Feb 11, 2026 · release notes

     Portkey Enterprise Gateway 2.2.2 adds Claude 4.6 support, native Responses API providers, and reasoning controls

     • Adds thinking: { type: 'adaptive' } with reasoning_effort for Claude 4.6
     • Adds output_config.format passthrough for structured outputsstructured outputsForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. in Anthropic/Bedrock/Azure
     • Adds cache_control preservation on content items and tools in the Responses API adapter
     • Adds native provider support for x-ai (Grok), Groq, OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code., and Azure AI Foundry in Responses API

     4 details in Everything

127. Data Service 1.5.0 Feb 4, 2026 · release notes

     Portkey Data Service adds dedicated batch pricing configurations via Gateway integration.

     └──▷ BREAKING ON UPGRADE

     • !Requires Enterprise Gateway version v2.1.0 or higher to support dedicated batch pricing.
128. Enterprise Gateway 2.1.0 Feb 4, 2026 · release notes

     Responses API now works with all providers via /v1/responses; batch pricing configs and Vertex cost attribution skip added

     • Adds support for the OpenAI Responses API (/v1/responses) with all 70+ providers, not just OpenAI and Azure OpenAI.
     • Adds vertex_skip_ptu_cost_attribution config key in Model Catalog settings to skip cost attribution for Vertex AI Provisioned Throughput deployments.
     • Adds dedicated batch pricing configurations via the Data Service, defaulting to 50% of standard pricing when specific pricing is unavailable.
     • Adds vertexSkipPtuCostAttribution field to the Virtual Key API for Vertex AI.

     4 details in Everything

129. Backend 1.8.0 Feb 4, 2026 · release notes

     Portkey Backend 1.8.0 introduces Gateway Deployments with RBACRBACRole-based access control, where permissions are assigned to roles such as reader or administrator and users receive those roles. It makes access easier to manage than individual permission lists., new CrowdStrike integration, and workspace manager API key creation.

     • Adds production and non-production deployment types with separate limits
     • Adds deployment slugs for human-readable identifiers in the Gateway Deployments system
     • Adds dedicated RBAC permissions (deployments:create, deployments:read, deployments:update, deployments:delete, deployments:list) for Gateway Deployments
     • Adds support for filtering deployments by type in the list endpoint
     • Adds CrowdStrike AIDR partner integration for scanning LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. inputs and outputs

     +4 more show less

     • Enhances Usage & Rate Limit Policies with new conditions (refer to Documentation)
     • Routes alert emails to workspace admins and managers in addition to organization admins
     • Enables Workspace managers to create API keys on behalf of other users by specifying a user_id
     • Adds optional skipPtuCostAttribution field for Vertex AI integrations and virtual keys

     9 details in Everything

130. Frontend 1.6.1 Feb 4, 2026 · release notes

     Frontend 1.6.1 adds multi-org SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management., workspace-level MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. access control, and saved filter search.

     • Adds SSO login support for multiple SSO enabled Organisations in the security settings UI
     • Adds Policy management UI Updates to support Organisation level multiple guardrails
     • Adds MCP capabilities support for the MCP Gateway and MCP access control support at Workspace level
     • Adds Search support for Saved Filters in the UI
     • Adds Logs Table column resize support

     5 details in Everything

131. February 2026 Feb 2026 · release notes

     Portkey adds gRPCgRPCAn open-source remote procedure call framework from Google that uses HTTP/2 and Protocol Buffers to let services call each other's functions across a network, giving tools low-latency, strongly typed inter-process communication. transport, universal Messages/Responses APIs, Zscaler Guard integration, and default config in JWTsJWTsJSON Web Tokens — an open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object. Cyber tools use them for stateless authentication without server-side session storage.

     └──▷ HOW TO FIND IT
     Install the official Portkey skills to teach Cursor or Copilot how to use the SDK.
     📍npx add-skill portkey-ai/skills
     • /v1/responses endpoint now routes to any provider (not just OpenAI).
     • /v1/messages adapter now works with non-Anthropic providers.
     • gRPC transport support added to the Gateway alongside HTTP/REST.
     • JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. defaults claim allows embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. default configs per-token.
     • npx add-skill portkey-ai/skills installs Portkey skills for AI coding assistants.

     +8 more show less

     • ZscalerAI Guard guardrail integration supports detection policies and DLPDLPData Loss Prevention: a category of controls that inspect and block sensitive data from leaving an organisation via email, web, or removable media, giving cyber tools a policy enforcement layer over data in motion and at rest. blocking.
     • New reasoning_effort parameter supported on Claude 4.6 and Together AI models.
     • output_config structured outputstructured outputForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. format available via Claude 4.6 adapter.
     • Veo model support on Vertex AI enables text-to-video generation.
     • Passthrough provider sends raw requests without Portkey parameter mapping.
     • stream_format: "sse" enables Server-Sent Events streaming for OpenAI/Azure TTS.
     • ZhipuAI cogview-4-250304 image generation supported.
     • media_resolution, vertex_skip_ptu_cost_attribution, and x-portkey-vertex-auth-type headers added for Vertex AI control.

     13 details in Everything

132. Enterprise Gateway 2.0.0 Jan 28, 2026 · release notes

     Enterprise Gateway 2.0.0 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. GA, unified rerank API, new budget condition keys, inline images plugin, and dynamic pricing.

     • MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. is now Generally Available with centralized server management, OAuth 2.1OAuth 2.1A consolidated draft of the OAuth authorization framework that merges OAuth 2.0 with its most important security BCP updates, dropping unsafe flows so tools can delegate user access without handling credentials directly./API key auth, workspace-level provisioning, and full observability
     • /rerank unified endpoint provides consistent rerankingrerankingA model that re-scores a first batch of search results against the query and reorders them, more accurately than the fast first-stage search. across Cohere, Jina, and Voyage AI without code changes
     • New condition keys for budget/rate limit policies: virtual_key, provider, config, prompt, and model (with wildcard support)
     • Inline Image URLs Plugin converts external image URLs to inline base64 data for VPCVPCA logically isolated virtual network inside a public cloud provider's infrastructure, letting tools deploy resources with private addressing, controlled routing, and firewall rules without managing physical hardware.-SC environments
     • MODEL_CONFIGS_PROXY_FETCH_ENABLED=ON environment variable enables dynamic model pricing fetch in Air-Gapped deployments via Backend service

     +1 more show less

     • GCP Workload Identity support for keyless authentication in GKEGKEGoogle Kubernetes Engine, Google's managed Kubernetes service, which lets cyber tools run containerized workloads without managing the underlying cluster infrastructure. environments accessing GCS and VertexAI

     6 details in Everything

133. Enterprise Gateway 1.17.15 Jan 23, 2026 · release notes

     Adds CrowdStrike AIDR partner plugin integration for LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. input/output scanning in Enterprise Gateway 1.17.15

134. Backend 1.7.3 Jan 21, 2026 · release notes

     Portkey Backend 1.7.3 adds tag support for integrations/providers, MCP OAuthMCP OAuthAn authentication layer for the Model Context Protocol that lets MCP clients obtain delegated access tokens so cyber tools can call MCP servers on a user's behalf without handling raw credentials. email/username fields, and JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. scopes.

     • Adds tags parameter support when creating or updating integrations/providers and allows filtering by tags in list endpoints
     • MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. OAuth token introspection endpoint now returns email and username fields for richer user context
     • Added read and list scopes to control plane resources for JWT-authenticated requests

     3 details in Everything

135. Enterprise Gateway 1.17.12 Jan 8, 2026 · release notes

     Portkey Enterprise Gateway 1.17.12 adds Responses API Hooks, Azure Shield Prompt/Protected Material guardrails, and sequential execution.

     • Apply input/output guardrails and custom webhooks to /v1/responses endpoint via new Responses API Hooks support
     • Detect jailbreak and prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. attacks using Shield Prompt Azure AI Content Safety integration
     • Detect copyrighted or protected text content in LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. outputs using Protected Material Azure AI Content Safety integration
     • Execute guardrails sequentially using the new sequential flag instead of parallel execution
     • Add real-time voice interaction support for xAI via Realtime Voice Agent API (Grok Voice API) with WebSocketWebSocketA protocol that upgrades an HTTP connection into a persistent, full-duplex channel, letting a server push data to a client without repeated polling. Cyber tools use it for real-time event streaming and live command output. connections

     +5 more show less

     • Add Google Maps grounding support for Gemini/Vertex AI models
     • Support new parameters for gpt-image-1: moderation, output_format, output_compression, background, partial_images, and stream
     • Specify custom authentication scopes for Azure OpenAI Entra IDEntra IDMicrosoft's cloud-based identity and access management service, formerly Azure Active Directory, used by cyber tools to authenticate users and enforce access policies in Microsoft 365 and Azure environments. auth using the x-portkey-azure-entra-scope header
     • Support output_expires_after parameter for batch creation, plus input_blob and output_folder for Azure OpenAI blob-based inputs
     • Enable Anthropic citations feature via native Messages API format on Bedrock /v1/messages route

     10 details in Everything

136. Backend 1.7.0 Jan 8, 2026 · release notes

     Portkey Backend v1.7.0 adds workspace deployment restrictions, SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. auto-provisioning, OCIOCIOpen Container Initiative standard, maintained by the Linux Foundation, defining the image format and runtime spec that container tools share so images built once run anywhere OCI-compliant. provider support, and new model configuration environment variables.

     • Adds workspace-based deployment restrictions to gateways for multi-tenant isolation
     • Adds endpoints to assign workspaces during gateway create/update operations
     • Adds a new API endpoint to retrieve entities (API keys, workspaces, users) associated with usage and rate limit policies
     • Enables automatic provisioning of first-time SSO users via OIDCOIDCOpenID Connect, a sign-in protocol built on OAuth 2.0. An application uses identity information from a trusted provider instead of maintaining a separate password for each user. or SAMLSAMLA standard for passing signed identity assertions between a sign-in provider and an application. It lets a user sign in through their organization's identity provider.
     • Adds Oracle Cloud Infrastructure (OCI) Generative AI as a supported provider

     +2 more show less

     • Adds MODEL_CONFIGS_PROXY_FETCH_ENABLED, MODEL_CONFIGS_PROXY_URL, MODEL_CONFIGS_PRICING_LOG_STORE_PATH, MODEL_CONFIGS_CAPABILITIES_LOG_STORE_PATH, MODEL_CONFIGS_CAPABILITIES_LOCAL_PATH, and MODEL_CONFIGS_PRICING_LOCAL_PATH environment variables for model config sourcing
     • Enables automatic reactivation of exhausted workspace budgets when credit limits are increased

     7 details in Everything

137. Frontend 1.6.0 Jan 7, 2026 · release notes

     Frontend 1.6.0 adds per-key expiration (next reset date) and model capability sync in UI.

     • Added configuration changes to sync with the models list and model capabilities specific to provider

     1 detail in Everything

138. Backend 1.6.2 Jan 7, 2026 · release notes

     Portkey Backend 1.6.2 adds auto-expiry for API keys and MySQL SSLSSLA cryptographic protocol for encrypting network connections, now deprecated in favor of TLS but whose name persists colloquially. Cyber tools reference it when handling HTTPS or other encrypted-channel inspection and termination. configuration via env vars.

     • Expired API keys are now automatically marked as expired by a background worker process
     • Added support for external MySQL SSL connection modes via the DB_SSL environment variable

     2 details in Everything

139. Frontend 1.5.7 Jan 6, 2026 · release notes

     Adds nested Prompt folders, Oracle Provider, sequential Guardrails execution, and request/response breakdown to the UI

     • Added nested Prompt folder support on UI
     • Added Oracle Provider support on UI
     • Added option to execute Guardrails sequentially
     • Added support for breakdown of LLM requests and responses
     • Added UI support for Retry Exhausted state for Logs

     5 details in Everything

140. January 2026 Jan 2026 · release notes

     Portkey introduces MCP GatewayMCP GatewayA server that sits between MCP clients and one or more Model Context Protocol backends, routing tool calls and aggregating context sources so a single endpoint exposes many AI-agent integrations at once. GA, Unified Rerank API, Azure Shield/Protected Material guardrails, and enhanced rate-limit policy matching.

     • MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Gateway is now generally available for unified MCP tool access with built-in auth support, central registry, RBACRBACRole-based access control, where permissions are assigned to roles such as reader or administrator and users receive those roles. It makes access easier to manage than individual permission lists., and full observability
     • Adds the sequential flag to enable sequential execution of guardrail checks instead of parallel
     • Integrates CrowdStrike AI Detection and Response as a new guardrail integration
     • Adds Azure Shield Prompt guardrail for jailbreak and prompt injectionprompt injectionAn attack where text the model reads, such as a web page, file or tool result, carries instructions that override the user's. It is the main security risk for agents with tools. detection in system prompts and user messages
     • Adds Azure Protected Material guardrail to scan LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. outputs for known copyrighted or protected text

     +4 more show less

     • Adds the /rerank endpoint for a unified rerankingrerankingA model that re-scores a first batch of search results against the query and reorders them, more accurately than the fast first-stage search. API supporting Cohere, Jina, and Voyage AI
     • Expands rate limit policy matching to include virtual_key, provider, config, prompt, and model (with wildcard support)
     • Adds support for OpenAI gpt-image-1 parameters including moderation, output_format, output_compression, background, partial_images, stream, and batch creation with output_expires_after
     • Enables Google Maps grounding to add map and location context to Gemini and Vertex AI responses

     9 details in Everything

141. Enterprise Gateway 1.17.11 Dec 23, 2025 · release notes

     Enterprise Gateway 1.17.11 adds ENABLE_PROMETHEUS env var to disable metrics collection.

142. Enterprise Gateway 1.17.9 Dec 18, 2025 · release notes

     Portkey Enterprise Gateway 1.17.9 adds Azure image edits and Gemini image config

     • Adds image_config parameter with aspect_ratio and image_size fields for Gemini/Vertex AI image generation

     1 detail in Everything

143. Backend 1.6.1 Dec 13, 2025 · release notes

     Portkey Backend v1.6.1 adds Guardrails config keys, new providers, and SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. parsing fixes

     • Adds guardrails configuration schema with tools (blockedTypes, allowedTypes, blockedFunctionNames, allowedFunctionNames) and params (blockedKeys, allowedKeys, values) sub-keys to control request filtering
     • Adds Javelin, Qualifire, and Walled as new plugin providers
     • Updates /spans endpoint aggregate metrics to include total tokens count

     3 details in Everything

144. Frontend 1.5.5 Dec 13, 2025 · release notes

     Portkey Frontend 1.5.5 adds embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG./completion support, prompt copy, and new guardrail/plugin UI

     • Added embedding text and completion support on getting started and model integrations
     • Added copy option on Prompt Playground on Each messages
     • Updated new Guardrails and Plugins support

     3 details in Everything

145. Enterprise Gateway 1.17.7 Dec 12, 2025 · release notes

     Portkey Enterprise Gateway v1.17.7 adds Blocked Tools guardrail, Redis ClusterRedis ClusterA distributed mode of Redis, the open-source in-memory data store, that automatically shards data across multiple nodes and keeps it available if some nodes fail, giving cyber tools low-latency shared state at scale. discovery, and semantic cache config env vars.

     • REDIS_CLUSTER_ENDPOINTS, REDIS_CLUSTER_DISCOVERY_URL, REDIS_CLUSTER_DISCOVERY_AUTH, and REDIS_CLUSTER_DISCOVERY_REFRESH_INTERVAL environment variables support dynamic Redis cluster discovery and static configuration
     • SEMANTIC_CACHE_EMBEDDING_DIMENSIONS environment variable allows setting custom embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. dimensions for the semantic cache (default: 1536)
     • New Blocked Tools guardrail plugin adds function, web_search_preview, web_search, file_search, code_interpreter, computer_use, and mcp tool type blocking via blocklists or allowlists

     3 details in Everything

146. Data Service 1.4.2 Dec 3, 2025 · release notes

     Adds file_upload_failed and batch_start_failed statuses for LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. gateway batch jobs.

     • file_upload_failed status indicates the batch job failed to upload the file to the provider after validation
     • batch_start_failed status indicates the file upload succeeded but the batch job submission to the provider failed

     2 details in Everything

147. Enterprise Gateway 1.17.5 Dec 1, 2025 · release notes

     Enterprise Gateway 1.17.5 adds SSRFSSRFServer-side request forgery, an attack that tricks a server into making requests chosen by an attacker. Those requests can reach internal services the attacker cannot access directly. protection, configurable vector similarity metrics, and provider auth via query params.

     • Added request validation to prevent Server-Side Request Forgery (SSRF) attacks via configuration URLs
     • Added configurable similarity metric support for Milvus and Pinecone vector storesvector storesA database built to store embeddings and find the nearest ones to a query quickly. It is the retrieval half of most RAG systems. in Semantic Cache (COSINE, L2, IP)
     • Added support for provider authentication via query parameters in the Realtime API (e.g., model=@provider/model)
     • Set span.kind = SERVER on exported OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. spans

     4 details in Everything

148. Enterprise Gateway 1.17.4 Nov 27, 2025 · release notes

     Portkey Enterprise Gateway 1.17.4 adds Anthropic Claude models and native messages endpoint via Azure AI Foundry

     • Adds Azure AI Foundry support for Anthropic Claude models (claude-opus-4-5-20251101, claude-sonnet-4-5-20250929, claude-haiku-4-5-20251001)
     • Enables native /messages endpoint support for Azure AI Foundry Anthropic providers, including extended thinkingextended thinkingIntermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens., prompt caching, and native streaming

     2 details in Everything

149. Enterprise Gateway 1.17.2 Nov 24, 2025 · release notes

     Adds cost calc for Gemini image models, thought_signature support, and Bedrock AWS region header

     • Adds cost calculation support for gemini-2.5-flash-image and gemini-3-pro-image-preview models
     • Adds support for the thought_signature parameter.
     • Added support for the x-portkey-aws-region header to set the region for Bedrock and SageMaker integration with serviceRole auth type

     3 details in Everything

150. Frontend 1.5.4 Nov 24, 2025 · release notes

     Frontend 1.5.4 adds SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. group workspace mapping, SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. confirmation modal, and AWS region config for service role auth.

     • SCIM group workspace mapping management available in UI to configure which SCIM groups map to which workspaces
     • Confirmation modal added before enabling the SSO Only Login flow to prevent accidental lockout
     • Support for specifying an AWS region when using service role-based authentication for Bedrock and Sagemaker

     3 details in Everything

151. Backend 1.6.0 Nov 24, 2025 · release notes

     Portkey Backend 1.6.0 adds SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. group-to-workspace mapping and AWS Bedrock region parameter support

     • Introduced flexible group-to-workspace mapping for SCIM integrations from Okta or Azure Entra
     • Added support for the aws_region parameter in the Bedrock integration for serviceRole auth type

     2 details in Everything

152. Enterprise Gateway 1.17.1 Nov 19, 2025 · release notes

     Portkey Enterprise Gateway 1.17.1 adds streaming hook results and F5 Guardrails plugin

     • Adds hook results in streaming responses for /chat/completions, /completions, /embeddings, and /messages endpoints to enable real-time content policy enforcement.
     • Introduces new F5 Guardrails partner plugin integration for content moderation and PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. detection/rediation.
     • Expands cost calculation support for image generation (TogetherAI), video generation (VertexAI Veo models, OpenAI/Azure Sora models).

     3 details in Everything

153. Backend 1.5.1 Nov 19, 2025 · release notes

     Portkey Backend v1.5.1 adds F5 Guardrails dependencies and log path support

     • Added backend dependencies for the new F5 Guardrails
     • Added support to read logs based on the path format identifier released in Gateway v1.17.0

     2 details in Everything

154. Frontend 1.5.1 Nov 18, 2025 · release notes

     Portkey Frontend 1.5.1 adds OpenAI gpt5.1 support, base64 image download in logs, and integration UI enhancements.

     • gpt5.1 models are now supported
     • Logs display any base64 URL with a Download option
     • Selected models prominently displayed at the top of the Update Integration Models page

     3 details in Everything

155. Frontend 1.5.0 Nov 17, 2025 · release notes

     Frontend 1.5.0 adds workspace member config management, trace/log UI categories, zhipu models, and non-root container security.

     • Non-root container image enforcing a non-root securityContext replaces the previous root-user image
     • Traces are highlighted by their category in the UI
     • Logs with Guardrails are improved for visibility in the UI
     • New models added to the zhipu provider integration

     4 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !The container image no longer runs as root; it enforces a non-root securityContext requiring a Helm repo upgrade (>app-1.4.0)
156. Backend 1.5.0 Nov 17, 2025 · release notes

     Portkey Backend 1.5.0 introduces usage limits and rate limit policy APIs with dynamic condition support.

     • Adds usage limits and rate limit policy APIs to allow organizations to apply flexible budget and rate limit controls based on dynamic conditions such as API keys, metadata, or workspace
     • Adds a new column in ClickHouse to store the log object path format identifier

     2 details in Everything

157. Data Service 1.4.0 Nov 17, 2025 · release notes

     Data Service 1.4.0 introduces non-root container enforcement and multi-provider fine-tuning cost logic.

     • Adds cost calculation logic for fine-tuning operations across OpenAI, Azure OpenAI, and Vertex AI

     1 detail in Everything

     └──▷ BREAKING ON UPGRADE

     • !The container image no longer runs as root; processes now run with a non-root UID, requiring Helm repo upgrade (>app-1.4.0)
158. Enterprise Gateway 1.17.0 Nov 17, 2025 · release notes

     Portkey Gateway 1.17.0 adds usage limits, OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. tracing, fine-tuning cost tracking, and hourly S3 logs

     • Adds LOG_STORE_FILE_PATH_FORMAT environment variable to toggle log storage between flat (v1) and time-hierarchical (v2) paths.
     • Introduces usage limits and rate limit policies based on API keys, metadata, workspace, or other dynamic conditions.
     • Adds support for OpenTelemetry W3C trace context headers (traceparent and baggage) to automatically map to internal tracing.
     • Adds cost calculation logic for fine-tuning operations across OpenAI, Azure OpenAI, and Vertex AI.

     4 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !Removed root user from container image. The container image no longer runs as root; it now runs processes with a non-root UID and enforces a non-root container securityContext.
159. Backend 1.4.0 Nov 12, 2025 · release notes

     Portkey Backend 1.4.0 introduces API-driven deployment management and customizable branding via environment variables, while enforcing non-root container execution.

     └──▷ BREAKING ON UPGRADE

     • !The container image no longer runs as root; processes now use a non-root UID and enforce a non-root securityContext, requiring Helm repo update.
160. Backend 1.3.3 Nov 7, 2025 · release notes

     v1.3.3: List traces API returns more details; JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. guardrail schema updated for new Gateway parameters

     • Updates JWT Guardrail schema to support new parameters from the latest Gateway build

     1 detail in Everything

161. Enterprise Gateway 1.16.6 Nov 6, 2025 · release notes

     Enterprise Gateway 1.16.6 adds llm_token_sum metric, Modal provider, JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. plugin claims injection, and skip.auth env var

     • Adds llm_token_sum Prometheus metric to track total LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. tokens consumed
     • Adds support for Modal as a new LLM provider
     • Adds claim extraction and injection into request context as headers in JWT Plugin
     • Adds flexible claim validation with multiple match types (exact, contains, containsAll, regex) in JWT Plugin
     • Adds token introspection endpoint validation as an alternative to JWKSJWKSA JSON Web Key Set, a published collection of cryptographic keys. Applications use it to find the public key needed to verify a token's signature. in JWT Plugin

     +1 more show less

     • Adds SKIP_DATAPLANE_CONFIG_CHECK environment variable to allow configs with multiple targets for unified batches file upload endpoints

     6 details in Everything

162. Frontend 1.4.9 Nov 6, 2025 · release notes

     Portkey Frontend v1.4.9 adds JSON modeJSON modeForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. for prompts, workspace search on MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. Consent, and Azure/OpenAI transcribe model support.

     • Adds JSON mode for prompt parameters to enable real-time updates and easier editing
     • Improves MCP Consent page to allow searching by workspaces
     • Supports transcribe models under Azure and OpenAI providers
     • Adds search functionality to the Analytics Filter API key field

     4 details in Everything

163. Enterprise Gateway 1.16.5 Nov 3, 2025 · release notes

     Portkey Enterprise Gateway 1.16.5 adds VertexAI Computer Use support, WebSocketWebSocketA protocol that upgrades an HTTP connection into a persistent, full-duplex channel, letting a server push data to a client without repeated polling. Cyber tools use it for real-time event streaming and live command output. proxy, and new CORSCORSA browser mechanism that lets a server declare which foreign origins may read its HTTP responses, enforced by the browser to block unauthorized cross-site requests from JavaScript. env vars

     • Adds anthropic-beta header support (also mapped to x-portkey-anthropic-beta) for VertexAI provider
     • Adds support for conversation and modalities parameters for OpenAI provider
     • Adds HTTP proxy support for WebSockets via the HTTPS_PROXY environment variable
     • Adds four CORS configuration environment variables: CORS_ALLOWED_ORIGINS, CORS_ALLOWED_METHODS, CORS_ALLOWED_HEADERS, and CORS_ALLOWED_EXPOSE_HEADERS

     4 details in Everything

164. Backend 1.3.2 Oct 29, 2025 · release notes

     Portkey Backend 1.3.2 adds Workspace Grouped Analytics API and Multi-Organization SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. support

     • Adds GET /workspaces/analytics API endpoint for workspace-grouped analytics with multi-dimensional grouping and cost-based filtering
     • Enables authentication via SSO across multiple organizations

     2 details in Everything

165. Enterprise Gateway 1.16.4 Oct 29, 2025 · release notes

     Enterprise Gateway 1.16.4 adds bearer tokenbearer tokenA credential that grants access to whoever presents it, without a separate proof of possession. Protecting the token matters because a stolen copy can be used by someone else. auth for AWS Bedrock and EKSEKSAmazon's managed Kubernetes service, run by AWS, that handles control-plane provisioning so teams can deploy containerized workloads without operating the underlying cluster infrastructure themselves. Pod Identity support.

     • Adds support for bearer tokensbearer tokensA credential that grants access to whoever presents it, without a separate proof of possession. Protecting the token matters because a stolen copy can be used by someone else. authentication in AWS Bedrock provider configuration
     • Adds EKS Pod Identity integration with IRSAIRSAAn AWS mechanism that lets a Kubernetes pod assume an IAM role via a service account, so tools running in EKS can call AWS APIs without embedding long-lived credentials. for EKS clusters

     2 details in Everything

166. Backend 1.3.1 Oct 23, 2025 · release notes

     Portkey Backend 1.3.1 introduces Add Prefix and Allowed Request Types guardrails.

     • Adds the Add Prefix guardrail to prepend configurable text to user input before model inference
     • Adds the Allowed Request Types guardrail to restrict processing to specific endpoints via allowlist or blocklist

     2 details in Everything

167. Enterprise Gateway 1.16.1 Oct 23, 2025 · release notes

     Adds configurable request prefixing and request type allowlisting/blocklisting in Enterprise Gateway v1.16.1

     • Adds the ability to set a configurable prefix on user input before it reaches the model
     • Adds control over Allowed Request Types via allowlist or blocklist filtering

     2 details in Everything

168. Enterprise Gateway 1.16.0 Oct 18, 2025 · release notes

     Enterprise Gateway 1.16 adds AWS Service Role auth, request timeout and NO_PROXY env vars, TLS for control plane, Bedrock global profile/name mapping.

     • Adds REQUEST_TIMEOUT environment variable to configure fetch timeouts for outbound LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. requests in milliseconds (default: 300000).
     • Adds NO_PROXY environment variable support to bypass outbound requests to specified hosts.
     • Adds TLS_KEY, TLS_CERT, and TLS_CA environment variables to configure TLS for control plane calls in air-gapped deployments.
     • Adds AWS Service Role authentication support for AWS Bedrock models, allowing the Gateway to use its own service role.
     • Adds global profile support for AWS Bedrock models.

     +1 more show less

     • Adds name field mapping for the /messages endpoint document object in AWS Bedrock.

     6 details in Everything

169. Backend 1.3.0 Oct 17, 2025 · release notes

     Portkey Backend 1.3.0 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. APIs, SCIMSCIMA standard for creating, updating and removing user accounts and groups across applications. It helps an identity provider keep access in sync when people join, change roles or leave. org management, workspace key limits, and Budget resets

     • Adds mcp endpoints for MCP Preview management via new APIs
     • Adds OAuth support to MCP Preview management
     • Adds configs_security_settings in organization settings to control view/edit access by role
     • Adds workspace-level user_api_key_limits to restrict max keys per workspace
     • Adds configurable default_expiry and max_expiry for user API key expiry

     +5 more show less

     • Adds new SCIM provisioning setting to enforce organization-only user management via SCIM while allowing workspace management via Portkey
     • Adds periodic budget_reset functionality in the update workspaces API
     • Adds strict name filter parameter to list workspaces API
     • Adds aws_service_auth as a new auth type for AWS Bedrock integrations
     • Adds setting to disable the Getting Started page

     10 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !API Key Management API updates list api-keys response to no longer return raw keys for other user API keys
170. Frontend 1.4.6 Oct 17, 2025 · release notes

     Frontend 1.4.6 adds per-key rate limits, user API key security settings, and new model support for MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface.

     • Added multiple rate limit support: now possible to set both cost-based and token-based rate limits on a single key
     • Introduced a security setting to restrict the maximum number of User API keys that can be created per user
     • Introduced a security setting to configure the default and maximum expiry of User API keys
     • Updated the MCP OAuthMCP OAuthAn authentication layer for the Model Context Protocol that lets MCP clients obtain delegated access tokens so cyber tools can call MCP servers on a user's behalf without handling raw credentials. consent screen to allow selection from eligible workspaces
     • Added better states for consent approval, rejection, and failures in MCP

     +6 more show less

     • Hide embedding modelsembedding modelsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. in prompt playground
     • Allow updating base models for custom models
     • Made alias a required field for Azure-OpenAI deployment configurations
     • Added gpt-5-codex, gpt-5-pro, and several other OpenAI and Azure-OpenAI models
     • Added new Together AI and Grok models
     • Added AWS Bedrock Titan EmbeddingEmbeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. and Titan v1 models

     11 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !alias is now a required field for Azure-OpenAI deployment configurations
171. Data Service 1.3.0 Oct 16, 2025 · release notes

     Data Service 1.3.0 adds KMSKMSA managed service that creates, stores, rotates and controls access to encryption keys, usually backed by HSMs. Applications ask it to encrypt or decrypt rather than holding keys themselves. key support for custom batch file output and batch operation tracking.

     • v1.3.0 supports KMS keys for encrypting custom batch file output
     • Adds batch operation tracking for provider batch requests sent from Gateway

     2 details in Everything

172. Enterprise Gateway 1.15.6 Oct 2, 2025 · release notes

     Portkey Enterprise Gateway 1.15.6 adds image editing pricing support, input_audio param for Vertex, and Bedrock batch pass-through

     • Supports pricing configuration for image editing models on openai, azure-openai, and azure-foundry providers
     • Adds input_audio parameter support for the vertex provider
     • Allows passing through parameters to the Bedrock batch create endpoint

     3 details in Everything

173. Frontend 1.4.3 Oct 2, 2025 · release notes

     Frontend adds Claude Sonnet 4.5 models for Anthropic, AWS Bedrock, and VertexAI base providers.

174. Enterprise Gateway 1.15.5 Sep 30, 2025 · release notes

     v1.15.5 enables HTTP protobuf transport for OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. via OTEL_EXPORTER_OTLP_PROTOCOL and fixes guardrail registration.

175. Enterprise Gateway 1.15.4 Sep 25, 2025 · release notes

     Portkey Enterprise Gateway 1.15.4 adds max_tool_calls, safety_identifier, and top_logprobs parameters for the /responses API.

     • Adds support for the max_tool_calls parameter on the /responses API endpoint
     • Adds support for the safety_identifier parameter on the /responses API endpoint
     • Adds support for the top_logprobs parameter on the /responses API endpoint

     3 details in Everything

176. Frontend 1.4.0 Sep 23, 2025 · release notes

     Frontend UI enhancements for MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. server management traceability and model catalog code generation.

177. Enterprise Gateway 1.15.3 Sep 19, 2025 · release notes

     Portkey Enterprise Gateway 1.15.3 adds APAC cross-region inference profiles for AWS Bedrock and streaming token calc for gpt-5

     • Adds support for APAC cross region inference profiles in the AWS Bedrock provider configuration

     1 detail in Everything

178. Enterprise Gateway 1.15.2 Sep 16, 2025 · release notes

     Portkey Enterprise Gateway 1.15.2 adds KMSKMSA managed service that creates, stores, rotates and controls access to encryption keys, usually backed by HSMs. Applications ask it to encrypt or decrypt rather than holding keys themselves. for Bedrock uploads and custom Entra scopes.

     • Adds KMS support for file uploads to bedrock/AWS
     • Supports custom scope for entra auth with deprecated azure serverless models
     • Enables custom header support for OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. Export of analytics data

     3 details in Everything

179. Enterprise Gateway 1.14.3 Aug 28, 2025 · release notes

     Portkey Enterprise Gateway 1.14.3 introduces a conditional new models endpoint with backward-compatible proxying logic.

     • Adds a models endpoint that activates only when the request lacks a provider, virtual_key, or config containing those fields

     1 detail in Everything

180. Enterprise Gateway 1.14.0 Aug 15, 2025 · release notes

     Portkey Enterprise Gateway 1.14.0 adds Vertex AI timeRangeFilter support and OpenAI GPT-5 parameters.

     • timeRangeFilter support for Google Search tool in Vertex AI provider updates
     • New parameters added for GPT-5 compatibility in OpenAI and Azure OpenAI providers
     • Return reasoning messages if returned by the model in OpenRouterOpenRouterA unified API gateway that routes requests to dozens of LLM providers under one endpoint, letting cyber tools swap or compare models without rewriting provider-specific integration code.

     3 details in Everything

181. Enterprise Gateway 1.13.2 Aug 12, 2025 · release notes

     Adds the /v1/models unified endpoint compliant with OpenAI API spec

182. Enterprise Gateway 1.13.1 Aug 6, 2025 · release notes

     Portkey Enterprise Gateway v1.13.1 adds granular request processing time analytics and optimizes OpenAI Chat Completions latency.

     • Adds new analytics data point for granular processing time for requests
     • Improves OpenAI Chat Completions by eliminating extra processing time

     2 details in Everything

183. Enterprise Gateway 1.13.0 Aug 4, 2025 · release notes

     Portkey Enterprise Gateway v1.13 adds a unified messages API following Anthropic's spec for Bedrock, Anthropic, and Vertex AI.

     • Adds unified messages API endpoint supporting Anthropic’s messages specification
     • Extends the new API to work with AWS Bedrock, Anthropic, and Vertex AI models

     2 details in Everything

184. Enterprise Gateway 1.12.0...1.12.5 Aug 2, 2025 · release notes

     Enterprise Gateway 1.12.x adds Model Catalog, workspace budget limits, per-strategy circuit breakers, and automatic user attribution.

     • Adds Model Catalog support with updated API specification for managing model providers
     • Enforces Workspace Budget Limits to control spend and rate limits at the workspace level
     • Introduces Circuit Breaker capability configured per-strategy within tool configurations

     3 details in Everything

185. Data Service 1.2.6 Jul 22, 2025 · release notes

     Data Service 1.2.6 adds Azure Redis cache with Entra/Managed Identity support and HTTPSHTTPSHTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate. proxy for inter-POD communication.

     • azure cache now supports Entra and Managed Identity auth modes
     • All external calls and HTTPs communication between PODs now route through an HTTPS Proxy
     • Custom logs can include virtual keys via request headers

     3 details in Everything

186. Enterprise Gateway 1.11.9 Jul 10, 2025 · release notes

     Portkey Enterprise Gateway 1.11.9 adds S3 Object Lock support, new provider parameters and unified finish reason handling

     • Adds LOG_STORE_OBJECT_LOCK_RETENTION_ENABLED environment variable to enable S3 Object Lock and Retention-enabled buckets for the log store
     • Unifies finish_reason values across Anthropic and Bedrock models, retaining provider-returned values when x-portkey-strict-openai-compliance is set to false
     • Adds support for cost calculation for fine-tuned models on Vertex AI
     • Adds support for the Computer Use tool for Anthropic models on AWS Bedrock
     • Handles backward compatibility for Titan G1 embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. model encoding_format parameter on AWS Bedrock

     +2 more show less

     • Forwards max_completion_tokens parameter as-is instead of mapping it to max_tokens on Azure AI
     • Adds support for the web_search_options parameter for chat completions endpoint on OpenAI and Azure OpenAI

     7 details in Everything

187. Enterprise Gateway 1.11.7 Jun 27, 2025 · release notes

     Adds Azure Managed Identity support for Azure Containers in Enterprise Gateway v1.11.7

     • Improves error logging for Azure Managed Identity errors

     1 detail in Everything

188. Enterprise Gateway 1.11.6 Jun 21, 2025 · release notes

     Groq service_tier support, Anthropic tool-result prompt caching, and webhook guardrail verdict fixes

     • Adds service_tier configuration option for the Groq provider to control service tier selection
     • Enables Anthropic prompt caching for tool results and tool usetool useA model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. calls in provider configurations

     2 details in Everything

189. Data Service 1.2.3 Jun 17, 2025 · release notes

     Portkey Data Service 1.2.3 adds configurable FINETUNE_STATUS_CHECK_INTERVAL and BATCH_STATUS_CHECK_INTERVAL environment variables for fine-tuning and batch processing

     • Adds FINETUNE_STATUS_CHECK_INTERVAL configuration to control provider fine-tuning status check frequency
     • Adds BATCH_STATUS_CHECK_INTERVAL configuration to control provider batch processing status check frequency

     2 details in Everything

190. Enterprise Gateway 1.11.4 Jun 17, 2025 · release notes

     Portkey Enterprise Gateway v1.11.4 adds Bedrock AssumedRole, Anthropic computer use tool, Azure Redis auth, HTTPSHTTPSHTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate. proxy, and more.

     • Adds AssumedRole support for Bedrock application inference profiles
     • Adds multimodalmultimodalAble to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. support for Cohere and Titan providers
     • Adds support for Bedrock createTranscription, createTranslation, imageGeneration, batch, and files endpoints via Azure Foundry
     • Adds Anthropic computer use tool support
     • Adds file_url and mime_type fields for file content parts in Anthropic requests

     +7 more show less

     • Adds Gemini/Vertex Thinking mode support in VertexAI
     • Adds Azure Redis authentication modes: EntraID and ManagedIdentity
     • Adds separate REDIS_USERNAME and REDIS_PASSWORD environment variables for Redis Cache auth
     • Adds azure-redis value for the CACHE_STORE environment variable to enable Azure Redis Cache
     • Adds AZURE_REDIS_AUTH_MODE, AZURE_REDIS_MANAGED_CLIENT_ID, AZURE_REDIS_ENTRA_CLIENT_ID, AZURE_REDIS_ENTRA_CLIENT_SECRET, and AZURE_REDIS_ENTRA_TENANT_ID environment variables for Azure Redis Cache configuration
     • Adds HTTPS proxy support via the HTTPS_PROXY environment variable for all external calls
     • Adds virtual key inclusion in custom logs when passed via headers

     12 details in Everything

191. Enterprise Gateway 1.11.1 Jun 2, 2025 · release notes

     Adds OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. traces collector endpoint, Data Plane log exports, and new provider parameters

     • Adds the /v1/otel/v1/traces endpoint to collect OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. spans as Portkey traces.
     • Enables log exports on the Data Plane, allowing logs to be exported without sending them to the Control Plane (requires Dataservice).
     • Adds support for background and service_tier parameters for OpenAI providers.
     • Adds support for custom hosts and private links for Azure Plugins.

     4 details in Everything

192. Data Service 1.2.2 May 31, 2025 · release notes

     Portkey Data Service 1.2.2 adds Prometheus counters for batch processing metrics.

     • Adds Prometheus metric support for batch processing visibility with eight new counters (batch_count, batch_cost, batch_input_tokens, batch_total_tokens, batch_process_time, batch_success_row_count, batch_failure_row_count, batch_row_count) and provider/type labels.

     1 detail in Everything

193. Enterprise Gateway 1.11.0 May 17, 2025 · release notes

     Adds mTLSmTLSMutual TLS, an encrypted connection where both client and server present certificates to prove their identity. It helps services verify who they are communicating with. support for Gateway-to-internal-service communication in Enterprise Gateway 1.11.0.

     • Adds mTLS support for Gateway to internal services in air-gapped deployments

     1 detail in Everything

194. Enterprise Gateway 1.10.23 May 15, 2025 · release notes

     Portkey Enterprise Gateway 1.10.23 adds dimensions param for Vertex multi-modalmulti-modalAble to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. and JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. plugin auth

     • Adds dimensions parameter support in multi-modal embeddings for Vertex provider
     • Adds JWT authentication capability to the Plugins module for runtime validation

     2 details in Everything

195. Enterprise Gateway 1.10.22 May 8, 2025 · release notes

     Portkey Enterprise Gateway 1.10.22 adds Anthropic PDF file support, new env vars for OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. config, and three new AI providers.

     • Adds file_url and mime_type parameters to request payloads for OpenAI-spec-compatible file content parts in Anthropic-based models
     • Adds OTEL_SERVICE_NAME and OTEL_RESOURCE_ATTRIBUTES environment variables for OpenTelemetryOpenTelemetryAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. resource configuration
     • Adds Ncompass, Lepton, and Snowflake Cortex as supported providers for chat completions and other endpoints

     3 details in Everything

196. Enterprise Gateway 1.10.20 Apr 23, 2025 · release notes

     Portkey Gateway 1.10.20 adds strict metadata enforcement and disables API key config overrides

     • Adds strict metadata logging hierarchy (Workspace Default > API Key Default > Incoming Request) with admin values preventing override by request-level fields
     • Adds ability to disable default config override for API keys, returning 400 error on conflict

     2 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !Request-level metadata fields can no longer override admin-configured values due to strict enforcement hierarchy
197. Data Service 1.1.10 Apr 21, 2025 · release notes

     Portkey Data Service v1.1.10 adds S3 encryption header support and environment variable secret injection.

     • Added support for passing encryption headers while uploading stream data to S3
     • Added support for both file path and direct value from environment variables for secrets like redis connection

     2 details in Everything

198. Enterprise Gateway 1.10.19 Apr 18, 2025 · release notes

     Enterprise Gateway adds OTelOTelAn open standard and set of tools for collecting traces, metrics and logs from applications. It lets services export observability data without tying instrumentation to one monitoring vendor. metrics push, Milvus semantic caching, Azure Guardrails, prompts render API, and Vertex AI dimensions support.

     • Adds OTEL_PUSH_ENABLED and OTEL_ENDPOINT environment variables to push ClickHouse analytics (traces/spans) to an OTel collector
     • Adds Milvus vector storevector storeA database built to store embeddings and find the nearest ones to a query quickly. It is the retrieval half of most RAG systems. support for semantic caching via the VECTOR_STORE, VECTOR_STORE_ADDRESS, VECTOR_STORE_API_KEY, and VECTOR_STORE_COLLECTION_NAME environment variables
     • Adds a /v1/prompts/:id/render endpoint to access prompts rendering capabilities within the Gateway
     • Adds PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. detection support using Azure Language Service for Azure Guardrails
     • Adds support for dimensions parameter in Vertex AI embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.

     5 details in Everything

199. Enterprise Gateway 1.10.17 Apr 11, 2025 · release notes

     Portkey Enterprise Gateway 1.10.17 adds Response API support, use_retry_after_header retry logic, and org-level cache TTLTTLTime to live, a limit on how long data remains valid or how far a packet can travel. Caches use a time limit; IP packets use a hop limit..

     • Adds end-to-end support for the OpenAI and Azure OpenAI Response API.
     • Implements caching for stream requests.
     • Introduces cost calculation for tools like web_search, file_search and code_execution.
     • Updates Azure Inference integration to directly accept endpoints from the Azure Foundry dashboard.
     • Adds new retry setting use_retry_after_header that uses provider x-retry-after or x-retry-after-ms headers for 429 wait times.

     +5 more show less

     • Allows configuring the default max cache TTL at the organisation level.
     • Adds support for logprobs and top_logprobs request parameters in Azure OpenAI.
     • Adds support for response_format and search_recency_filter request parameters in Perplexity.
     • Handles empty assistant tools messages containing only newline characters (\n\n) in AWS Bedrock.
     • Populates the model field in /chat/completions API responses when providers do not natively return it.

     10 details in Everything

200. Enterprise Gateway 1.10.16 Apr 9, 2025 · release notes

     Enterprise Gateway v1.10.16 enhances the unified batch API's resilience and error visibility by adding error_file_id for OpenAI/Azure upload failures and improving Bedrock error handling.

     • Returns error_file_id in the batch output payload when file uploads fail for OpenAI and Azure OpenAI providers
     • Handles file upload failures for AWS Bedrock during unified batch processing

     2 details in Everything

201. Enterprise Gateway 1.10.15 Apr 2, 2025 · release notes

     Enterprise Gateway 1.10.15 adds large file upload support and custom mime-type configuration for providers.

     • Adds the ability to pass custom mime_type values in the request body for file uploads to Providers and Data Service
     • Enables uploading large files to Providers and Data Service

     2 details in Everything

202. Enterprise Gateway 1.10.14 Mar 28, 2025 · release notes

     Portkey v1.10.14 adds org/workspace guardrails, Fireworks support, logging updates, and new env var.

     • Enforces guardrails at the organisation and workspace levels on all requests
     • Extends unified finetuning APIs to support the Fireworks provider
     • Adds cost calculation for Perplexity search and Gemini grounding
     • Preserves custom_id in VertexAI batch output
     • Logs non-OpenAI compliant fields (e.g. groundingMetadata, citations) in stream responses

     +2 more show less

     • Adds support for logprobs and top_logprobs parameters in Fireworks provider requests
     • Introduces the AWS_ENDPOINT_DOMAIN environment variable to override the default AWS endpoint domain

     7 details in Everything

203. Data Service 1.1.5 Mar 21, 2025 · release notes

     Portkey Data Service v1.1.5 adds custom batch output paths and raises limits to 500k lines / 5MB chunks.

     • Increased maximum lines for custom batches to 500k
     • Raised chunk size limit to 5MB

     2 details in Everything

204. Data Service 1.1.4 Mar 20, 2025 · release notes

     Portkey Data Service v1.1.4 adds Vertex batch embeddingsembeddingsLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG., model logging, custom batch outputs, and TLS via env vars

     • Added TLS_KEY_PATH and TLS_CERT_PATH environment variables to enable internal pod-to-pod HTTPSHTTPSHTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate. communication by mounting a volume with certificate and key files.
     • Added support for Vertex batch embeddings as a new embeddingembeddingLists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. provider capability.
     • Includes model information in log objects, enabling visibility into which model was used for each embedding request.
     • Implemented custom batch processing output generations, allowing users to define specific output formats for batch jobs.

     4 details in Everything

205. Enterprise Gateway 1.10.12 Mar 18, 2025 · release notes

     Portkey Enterprise Gateway 1.10.12 adds Anthropic Extended Reasoning support, internal POD HTTPSHTTPSHTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate. via env vars, and new Prometheus metrics

     • Adds unified API signature for single-turn and multi-turn conversations with Anthropic Extended Reasoning across Anthropic, AWS Bedrock, and VertexAI.
     • New Prometheus metric llm_last_byte_diff_duration_milliseconds tracks LLMLLMA large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. last byte latency for chunked JSON responses.
     • All Prometheus metrics now include a stream label with possible values of 0 or 1.

     3 details in Everything

206. Enterprise Gateway 1.10.11 Mar 6, 2025 · release notes

     Portkey Enterprise Gateway 1.10.11 adds S3 encryption key support, webM mimeType, Azure drop param, and streaming usage mapping

     • Added support for specifying an encryption key when uploading files to S3 for AWS Bedrock providers
     • Added webm mimeType support in VertexAI batches and fine-tune APIs
     • Changed Azure Inference provider to use extra-parameters: drop instead of ignore for extra parameter handling
     • Mapped usage object for streaming responses in Openrouter

     4 details in Everything

207. Enterprise Gateway 1.10.10 Feb 27, 2025 · release notes

     Portkey Enterprise Gateway 1.10.10 adds VertexAI finetuning APIs and body params in conditional router

     • Unified Finetuning APIs now support VertexAI for file uploads and transformations.
     • Adds params body field support to conditional router queries, enabling routing based on request body fields.

     2 details in Everything

208. Data Service 1.1.2 Feb 27, 2025 · release notes

     Portkey Data Service adds Vertex provider options for batches.

     • Added support for Vertex provider options for batches.
     • Implemented cluster mode Redis for queues.

     2 details in Everything

209. Data Service 1.1.1 Feb 20, 2025 · release notes

     Data Service 1.1.1 adds S3 encryption headers, SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. upload support, and log export filtering with limits.

     • Adds S3 encryption header support for finetunes and batches
     • Implements SSE file upload support
     • Made S3 bucket optional for Bedrock batches
     • Adds filtering for log exports
     • Implemented end limit for log export records

     5 details in Everything

210. Enterprise Gateway 1.10.8 Feb 13, 2025 · release notes

     Portkey Enterprise Gateway 1.10.8 adds mutator webhooks and configurable guardrail timeouts

     • Adds mutation capability to the existing webhook plugin via Mutator Webhooks for BYO-PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. redaction guardrails
     • Introduces timeout parameter for all guardrails that make internal fetch calls, with a default of 5 seconds
     • Enables storing timeout values in the control plane when creating/updating a Guardrail on UI

     3 details in Everything

211. Data Service 1.1.0 Feb 10, 2025 · release notes

     Data Service 1.1.0 adds Bull Board integration for job queue visualization

     • Adds Bull Board integration to visualize job queues and their status.
     • Adds Prometheus metrics for batch jobs and fine-tuning operations.

     2 details in Everything

212. Enterprise Gateway 1.10.6 Feb 7, 2025 · release notes

     Portkey Enterprise Gateway 1.10.6 adds AzureOpenAI batching via unified APIs and Deepseek model support.

     • Extends the Unified Batches APIs to support AzureOpenAI batching.
     • Adds support for Deepseek models across Fireworks, Groq and Together providers.

     2 details in Everything

213. Enterprise Gateway 1.10.5 Feb 6, 2025 · release notes

     Portkey Enterprise Gateway 1.10.5 adds JWTJWTAn open standard (RFC 7519) for encoding claims as a signed, optionally encrypted JSON object passed between parties. Cyber tools use it to verify identity and permissions without querying a central session store. auth and VertexAI batch support

     • Adds JWT Auth support with configurable JWKS endpoint or JWKS JSON for authentication and authorization

     1 detail in Everything

214. Enterprise Gateway 1.10.4 Jan 29, 2025 · release notes

     Portkey Enterprise Gateway 1.10.4 adds PIIPIIPersonally Identifiable Information — data that can identify a specific individual, such as names, email addresses, or SSNs. Cyber tools flag or redact it to limit exposure and meet privacy regulations. Redaction Guardrails, Replicate proxy support, and request-level metadata logging.

     • Adds PII Redaction Guardrails through Portkey Managed, AWS Bedrock, Pangea, Patronus, and Promptfoo providers with a transformed flag in the guardrail result object
     • Adds Replicate provider support for proxy (passthrough) requests
     • Logs workspace metadata at the individual request level

     3 details in Everything

215. Enterprise Gateway 1.10.3 Jan 23, 2025 · release notes

     Azure OpenAI finetuning, AWS Bedrock guardrails, Milvus passthrough, Prometheus metric updates

     • Extended unified finetuning APIs to support the AzureOpenAI provider
     • Added AWS Bedrock Guardrails support for request/response checks
     • Allowed configuring custom host and custom headers directly in virtual keys for custom models/providers
     • Renamed Prometheus label llm_request_duration_seconds to llm_request_duration_milliseconds and added new metric portkey_request_duration_milliseconds
     • Supported Milvus DB as a passthrough provider

     +2 more show less

     • Added logprobs and top_logprobs support for VertexAI and Google providers compatible with OpenAI format
     • Added support for experimental Gemini Thinking Models in VertexAI/Google provider

     7 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !The Prometheus label llm_request_duration_seconds is renamed to llm_request_duration_milliseconds, which may break dashboards or alerts relying on the old label name.
216. Enterprise Gateway 1.10.1 Jan 13, 2025 · release notes

     Portkey Enterprise Gateway 1.10.1 adds unified finetune APIs and code detection guardrail updates.

     • Adds unified finetune APIs for OpenAI, AzureOpenAI, Bedrock, and Fireworks
     • Adds checks for verbose identifiers to detect Python and JavaScript markdown code blocks in Code Detection Guardrails

     2 details in Everything

217. Enterprise Gateway 1.10.0 Jan 3, 2025 · release notes

     Portkey Enterprise Gateway 1.10.0 adds unified batch APIs, new partner guardrails, VertexAI self-deployed model support, and configurable guardrail fields.

     • Adds input_guardrails and output_guardrails config fields accepting arrays of slugs for shorthand guardrail application
     • Adds explanation property to guardrail responses clarifying pass/fail reasons for default checks
     • Adds unified batching APIs for OpenAI, AWS Bedrock, and Cohere providers
     • Adds support for self-deployed VertexAI models (Endpoints) including Vertex-Huggingface configurations
     • Adds new partner guardrails for Mistral mistral.moderateContent and Pangea pange.textGuard

     +2 more show less

     • Enables mapping of OpenAI developer role to system role on non-OpenAI/AzureOpenAI providers
     • Reduces Docker image size from ~275MB to ~75MB through build process updates

     7 details in Everything

     └──▷ BREAKING ON UPGRADE

     • !Removed unsupported stream parameter from Bedrock-Cohere integration
218. Enterprise Gateway 1.9.5 Dec 17, 2024 · release notes

     Portkey Enterprise Gateway 1.9.5 adds Gemini Grounding Mode support in OpenAI compatible format

219. Enterprise Gateway 1.9.4 Dec 11, 2024 · release notes

     Portkey Gateway v1.9.4 adds comprehensive request/response logging and standardizes Prometheus labels.

     • Adds Prometheus metric labels for method, route, code, custom_labels, provider, model, and source
     • Adds support for the tools capability in Ollama and Groq providers

     2 details in Everything

220. Enterprise Gateway 1.9.3 Dec 6, 2024 · release notes

     Portkey Enterprise Gateway 1.9.3 adds S3_CUSTOM log stores and AWS Sagemaker passthrough support.

     • Adds LOG_STORE_BASEPATH config key to support the new S3_CUSTOM LOG_STORE type for arbitrary S3-compatibleS3-compatibleAn object-storage API that mirrors Amazon S3's HTTP interface, so any client or cyber tool written for AWS S3 works unchanged against any compliant storage backend. storage providers
     • Adds support for passing requests through to AWS SageMaker models as a new provider

     2 details in Everything

221. Enterprise Gateway 1.9.2 Nov 29, 2024 · release notes

     Enterprise Gateway 1.9.2 adds prompt label cache invalidation, S3 log sub-paths, and Proxy virtual key support.

     • Add prompt_labels support for cache key invalidation on control plane updates
     • Allow sub-paths in S3 bucket names for logs

     2 details in Everything

222. Enterprise Gateway 1.9.1 Nov 25, 2024 · release notes

     Enterprise Gateway 1.9.1 adds new parameters for Realtime APIs and fixes Bedrock response formatting.

     • Adds prediction, store, metadata, audio and modalities parameters support for OpenAI and AzureOpenAI Realtime APIs
     • Retains the original close reason from the provider in the socket close event for Realtime APIs

     2 details in Everything

223. Enterprise Gateway 1.9.0 Nov 20, 2024 · release notes

     Enterprise Gateway v1.9.0 adds dynamic budgets, auto-expiry for keys, and Realtime API support for OpenAI/Azure.

     • Adds Realtime APIs integration capability for OpenAI and AzureOpenAI providers

     1 detail in Everything

224. Enterprise Gateway 1.8.1 Oct 30, 2024 · release notes

     Adds control_plane log/analytics store option to reduce ClickHouse dependency

     • Adds LOG_STORE and ANALYTICS_STORE environment variable options with value control_plane to route logs and analytics without a ClickHouse connection.
     • Introduces the new internal log store backend named control_plane.

     2 details in Everything

225. Enterprise Gateway 1.8.0 Oct 25, 2024 · release notes

     Enterprise Gateway 1.8.0 adds Bedrock Converse API, Vertex Imagen, Stable Diffusion v2, Pydantic SDK structured outputsstructured outputsForcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed., Lambda provider, and prompt cache pricing.

     • Bedrock chat completions now use the Converse API instead of /chat/completions, enabling tool calls and vision for supported models while removing manual chat templating logic.
     • Adds Vertex Imagen image generation support via the Gateway.
     • Adds support for Stable Image Ultra, Core, 3.0, and 3.5 models through stability-ai and bedrock providers, handling internal API signature transformations between v1 and v2.
     • Adds Pydantic SDK integration for structured outputs in GoogleAI and VertexAI providers by dereferencing JSON schemas to ensure Google compatibility.
     • Adds support for prompt caching pricing calculations for OpenAI and AzureOpenAI models that require it.

     +1 more show less

     • Adds lambda provider with support for chat completions and completions endpoints.

     6 details in Everything

226. Data Service 1.0.3 Nov 28, 2023 · release notes

     Portkey Data Service 1.0.3 adds vision fine-tuning validation and S3 storage for fine-tunes

     • Adds support for vision fine-tuning validation for OpenAI.
     • Implements S3 bucket support to store fine-tunes.

     2 details in Everything