Hermes

No entries match these filters. Clear filters

1. v2026.9.11 Sep 11, 2026 · release notes

   Hermes Agent v2026.9.11 adds password-blind credential vault, plugin catalog UI, and guided first launch.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v2026.9.11 https://github.com/NousResearch/hermes-agent.git
   # already have the repo? check out this version:
   $ git checkout v2026.9.11

   • Adds a single Plugins page in the Desktop UI managing agent and desktop plugins, install, catalog, and per-commit SHA-pinning
   • Exposes connector tools (Gmail, Linear, Notion) through searchable tool_search and supports the Radio SDK as an opt-in plugin

   2 details in Everything

2. v2026.8.31 Aug 31, 2026 · release notes

   Hermes v2026.8.31 ships Bot Mode with CLI peer-to-peer messaging via hermes peer, cron continuity, and live browser control.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v2026.8.31 https://github.com/NousResearch/hermes-agent.git
   # already have the repo? check out this version:
   $ git checkout v2026.8.31

   └──▷ USE IT
   Send a message from one bot to another by handle to transfer findings.
   shell copy

   hermes peer @coding_bot "Here are the vulnerability results: ..."

   Configure a scheduled cron job to maintain state and deduplicate output across runs.
   shell copy

   hermes cron add --name 'monitor' --schedule '0 9 * * *' --continuity=true

   • Adds hermes peer subcommand for bot-to-bot DMs by handle across profiles and gateways
   • Bundles Bot Mode into the desktop app, enabling named agent rosters and shared group chats
   • Gains persistent memory and scratchpad support for cron jobs via continuity=true and monitor-mode deduplication
   • Consolidates 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 into a unified desktop page with health checks, cost overlays, and hermes:// deep links for installation
   • Adds Ctrl+P fuzzy command palette and /model, /status CLI filters in the desktop app

   +2 more show less

   • Enables the agent to navigate, click, and read the desktop's in-app browser directly
   • Introduces model_overrides config key to patch context windowscontext windowsThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. and pricing per model

   7 details in Everything

3. v2026.8.27 Aug 27, 2026 · release notes

   Hermes v2026.8.27 adds consent-gated profile browsing, OS-keychain secret encryption, multi-query tool_search and 50+ 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. servers

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v2026.8.27 https://github.com/NousResearch/hermes-agent.git
   # already have the repo? check out this version:
   $ git checkout v2026.8.27

   • Consent-gated real-profile browsing using the default Chromium profile with a Windows close-with-approval flow
   • Opt-in OS-keychain encryption for stored secrets that removes per-launch macOS Keychain prompts
   • tool_search now supports multi-query inputs with stemming

   3 details in Everything

4. docs update Aug 18, 2026 · issue 049

   Hermes Agent launches with CLI + desktop install, multi-platform messaging gateway, Tool Gateway, and full configuration surface.

   └──▷ TRY IT
   Fastest onboarding path: authenticate with Nous Portal and enable the full Tool Gateway (web search, image generation, TTS, cloud browser) in one step.
   copy

   $ hermes setup --portal

   Install on a headless service account that has no sudo access and does not need browser automation.
   copy

   $ curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash -s -- --skip-browser

   Diagnose a broken or post-upgrade setup to see exactly what is missing and verify the detected install method.
   copy

   $ hermes doctor

   • Adds hermes setup --portal to log in to Nous Portal, set Nous as the 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, and enable the Tool Gateway (web search, image generation, TTS, cloud browser) in a single command.
   • Adds hermes model command to choose and configure the LLM provider and model.
   • Adds hermes tools command to configure which tools are enabled.
   • Adds hermes gateway setup command to connect messaging platforms (Telegram, Discord, Slack, and others).
   • Adds hermes config set and hermes config get commands to read and write individual configuration values.

   +12 more show less

   • Adds hermes setup wizard command to configure all settings at once.
   • Adds hermes desktop command to install and launch the Hermes Desktop application after a CLI-only install.
   • Adds hermes import for restoring a full backup and hermes profile import for importing a single agent profile when migrating to a new machine.
   • Adds hermes doctor diagnostics command that reports what is missing, how to fix it, and surfaces the detected install method.
   • Adds hermes config check and hermes config migrate commands for detecting and migrating missing or stale configuration after updates.
   • Adds hermes update command that auto-detects the install path (git installer, Docker, or NixOS) and prints the matching update command.
   • Supports --skip-browser installer flag to bypass the Playwright/Chromium installation step for headless or browser-automation-free deployments.
   • Supports --skip-computer-use installer flag to opt out of pre-installing cua-driver (the Computer Use toolset) during installation.
   • Supports root-mode FHS install layout (/usr/local/lib/hermes-agent/, /usr/local/bin/hermes) for shared-machine or system-service deployments, with per-user config still stored under ~/.hermes/ or $HERMES_HOME.
   • Supports non-sudo / unprivileged service-user installs with graceful degradation when sudo is unavailable — Chromium is installed into the user's local Playwright cache and the required admin command is printed.
   • Includes Tool Gateway giving access to web search, image generation, TTS, and a cloud browser under a single Nous Portal subscription.
   • Installer automatically provisions all dependencies — Python 3.11 (via uv), Node.js v22, ripgrep, and ffmpeg — with no manual prerequisite installation required.

   17 details in Everything

5. docs update Aug 18, 2026 · issue 049

   Hermes adds NVIDIA SkillEvaluator advisory scanning and project-local skills with per-repo trust controls.

   └──▷ TRY IT
   Install the NVIDIA scanner binaries so Hermes runs Tier 1 advisory checks on every hub install.
   copy

   $ uv tool install --python 3.13 "skillevaluator @ git+https://github.com/NVIDIA/SkillEvaluator.git@v0.1.0"
   uv tool install "git+https://github.com/NVIDIA/SkillSpector.git@v2.9.5"

   Trust a cloned repo so its .hermes/skills/ and .agents/skills/ project skills are loaded in sessions started inside it.
   copy

   $ cd ~/myproject && hermes skills trust

   Disable project-skill discovery globally so Hermes never scans repos or shows trust notices.
   ~/.hermes/config.yaml copy

   skills.project_discovery: false

   • Advisory scan findings are printed with file and line before install confirmation; credential-class findings (private keys, cloud access keys, tokens, credentialed connection strings) are highlighted in red; 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. findings are informational and never block install.
   • The dashboard's Browse-hub scan button now returns advisory scan data in a tier1 field alongside the built-in scanner verdict.
   • Adds project-local skills loaded from <project-root>/.hermes/skills/ (Hermes-native) or <project-root>/.agents/skills/ (cross-tool convention) when Hermes is launched inside a git checkout.
   • Adds hermes skills trust and hermes skills untrust subcommands to grant or revoke trust for a repo's project skills; trusted roots are stored in skills.trusted_project_dirs.
   • Adds skills.project_discovery: false config key to disable project-skill scanning and notices entirely.

   +4 more show less

   • Project skills occupy the highest-precedence tier (project → local (~/.hermes/skills/) → external_dirs), overriding same-named profile or bundled skills within their repo; project skills are tagged [project] in the skill index.
   • Every project skill is scanned with the same security scanner used for Skills Hub installs before entering the index; scan results are content-hash cached under ~/.hermes/cache/project_skill_scans/ and re-run automatically on content change.
   • Skills with a malicious verdict (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. directives, credential-exfiltration commands, hidden-text tricks) are quarantined and excluded from the skill index, slash commands, and refuse to load by name with an explanatory error.
   • Non-interactive surfaces (cron, API, ACP) inherit the interactive trust decision from the surface's working directory (workdir) without prompting or auto-trusting.

   9 details in Everything

6. docs update Aug 18, 2026 · issue 049

   Hermes adds hermes peer subcommand for cross-machine bot-to-bot messaging via registered peer gateways.

   • Adds hermes peer add <name> --url <URL> [--key <KEY>] [--note TEXT] to register or update a remote Hermes gateway as a peer, storing the key as HERMES_PEER_<NAME>_KEY.
   • Adds hermes peer list to display all registered peers and whether each has an API key configured.
   • Adds hermes peer dm <peer>[/<agent>] [message] to message a remote agent's canonical Bot Chat over a peer gateway and print the reply to stdout; <peer> targets the gateway's main agent, <peer>/<agent> targets a named profile routed via its /p/<profile>/ mirror.
   • Adds hermes peer remove <peer> to remove a peer from the registry (key entry is preserved).
   • When at least one peer is registered, the agent.bot_mode_protocol taught to every canonical Bot Chat automatically includes the peer roster and the hermes peer dm pattern, enabling agents to discover cross-machine teammates without manual SOUL edits.

   5 details in Everything

7. v2026.8.18 Aug 18, 2026 · issue 049

   Hermes v2026.8.18 adds Browser Use mode, hermes peer, Meta Muse Spark provider, per-plugin durable storage, and inline message components.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v2026.8.18 https://github.com/NousResearch/hermes-agent.git
   # already have the repo? check out this version:
   $ git checkout v2026.8.18

   • Adds hermes peer subcommand for bot-to-bot DMs across machines and gateways.
   • Adds Meta Model API (Muse Spark) as a bundled provider plugin, with Meta Muse Spark 1.2 available in 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. curated picker.
   • Adds per-plugin durable data directory that survives plugin updates and removal.
   • Enables plugins to render inline components inside assistant messages via ::name{...} directives, including ::preview which renders pages live inside the message.
   • Adds NVIDIA SkillEvaluator Tier 1 advisory scanning on skill installs, covering license and security checks.

   +5 more show less

   • Adds project-skill quarantine with non-interactive trust inheritance.
   • Adds cross-connection Bot Mode — Create-on picker and multi-machine group chats — plus disband (delete) for group chats.
   • Adds Browser Use mode A/B benchmark (feat(evals): add the Browser Use mode A/B benchmark) to the evals suite.
   • Widget clicks now reach the agent as hidden user turns, making widget interaction a first-class agentic response path.
   • Adds an out-of-sync warning in About when the app build diverges from the runtime, with a direct link to the installer.

   10 details in Everything

8. v2026.8.16.2 Aug 17, 2026 · issue 049

   Hermes v2026.8.16.2 adds Bot Mode plugin, remote gateway headers, Cua Driver 0.20 support, and /rollback --all flag.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v2026.8.16.2 https://github.com/NousResearch/hermes-agent.git
   # already have the repo? check out this version:
   $ git checkout v2026.8.16.2

   • Bundles Bot Mode (hermes-bots) as a built-in, default-on plugin with the core teammate protocol for Bot Chat sessions.
   • Adds support for remote gateway headers, carried through the connections registry, test probes, and Settings UI.
   • Supports Cua Driver 0.20 runtime contracts for computer-use, with auto-repair of installed drivers that fail the runtime contract.
   • Raises Codex OAuth context windowcontext windowThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. to 900K tokens for the gpt-5.6 family and gpt-5.4 (subscription 1M rollout), and to a live-verified 350K for those same models.
   • Adds connection-aware plugin routing in the desktop app, scoping plugin socket connections to the active profile.

   +3 more show less

   • Adds a status bar reconnect action for offline gateways, with self-healing for dropped SSH/HTTP registered remote connections.
   • Adds context-manager protocol support on SessionDB for safer session state management.
   • Supports UTF-16 file reads and Gemini 3 tool-call ID preservation.

   8 details in Everything

9. docs update Aug 17, 2026 · issue 048

   Hermes adds AGENTS.override.md for personal, per-directory overrides of committed AGENTS.md instructions.

   └──▷ TRY IT
   Keep personal agent instructions that differ from the team's committed AGENTS.md without touching the tracked file.
   copy

   $ echo 'Always respond in Spanish.' > AGENTS.override.md
   echo 'AGENTS.override.md' >> .gitignore

10. docs update Aug 17, 2026 · issue 048

    Hermes 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. suggestion metadata, invisible Unicode TAG stripping, and selective _meta passthrough for tool results.

    • Adds suggest block support in MCP server manifests, with keywords and hosts lists that trigger one-click Add <server> pills in the Desktop app composer when a matching word or hostname is detected in a draft.
    • Strips invisible Unicode TAG characters (U+E0000–U+E007F range) from MCP tool results, resource content, and tool descriptions to block 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. smuggling, while preserving legitimate emoji tag sequences such as regional flags.
    • Surfaces vendor _meta mappings (e.g. com.example/handoff) from MCP tool results to the model, while dropping protocol-reserved keys under modelcontextprotocol and mcp prefixes (e.g. modelcontextprotocol.io/..., tools.mcp.com/...); omits _meta entirely when no model-facing keys remain.
    • GitHub is intentionally excluded from the MCP hosted catalog; Desktop instead offers the github-auth skill when gh is not signed in, routing through bundled github/* skills driving the gh CLI.

    4 details in Everything

11. docs update Aug 17, 2026 · issue 048

    Hermes agent now annotates signal-killed terminal exits and auto-transcodes UTF-16 files to UTF-8.

    • Terminal commands killed by a signal now carry a human-readable annotation in the result — e.g. exit code -9/137 becomes 'terminated by signal 9: SIGKILL — often the kernel OOMOOMAn out-of-memory condition: a program requests more memory than the system or device can provide. The allocation fails or the operating system may stop the process. killer on memory exhaustion, or an explicit kill -9'; segfaults, aborts, SIGTERM, broken pipes, and CPUCPUThe main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware./file-size limit signals are labeled the same way.
    • UTF-16 files (detected via BOMBOMA Bill of Materials listing every component, library, and dependency in a software project; cyber tools use it to track what is in a codebase and flag known-vulnerable components quickly. or byte-pattern heuristic, either endianness — covering Windows Notepad files and PowerShell redirects) are now transcoded to UTF-8 for display instead of being refused as binary; the result includes a hint disclosing the conversion, and edits via write_file re-encode as UTF-8; files over 10 MB and genuinely binary files still receive the binary-file refusal.

    2 details in Everything

12. docs update Aug 17, 2026 · issue 048

    Hermes adds tail_mode compaction control, delegation.* subagent routing, and new CLI multiline/scrollback config keys.

    ▶ ⤢
    • Adds tail_mode config key with legacy (0.20× window verbatim tail) and lean (clamped 2.5% tail, 10K–25K tokens, with digests, anchor index, and session_search recovery pointers) modes — lean yields ~3× fewer retained tokens after compaction.
    • Adds delegation.* config section for subagent model routing: the entry key controls which provider, model, and credentials subagents inherit, with inherit meaning 'use the parent agent's settings'; persists to delegation.* rather than auxiliary.* because subagents are full child agents.
    • Adds cli_multiline_shortcuts config key to enable Ctrl+J, \ + Enter, and Shift+Enter as newline-insert shortcuts in the CLI (false restores legacy Ctrl+J submit behavior).
    • Adds cli_rebuild_scrollback_on_redraw config key to issue a CSICSIAn ambiguous acronym spanning Container Storage Interface, Crime Scene Investigation, and other expansions — without more context about which domain this tool operates in, a reliable definition cannot be given. 3J wipe of terminal scrollback on /redraw, Ctrl+L, or resize recovery — useful when a terminal or tmux stack stamps stale prompt chrome into scrollback on maximize/restore.

    4 details in Everything

13. docs update Aug 17, 2026 · issue 048

    Hermes adds CLI health checks, macOS permission management, and an interactive session picker with search and resume.

    • Adds doctor [--include CHECK] [--skip CHECK] [--json] command to run cua-driver's health report and display platform checks.
    • Adds permissions status [--json] command to report macOS Accessibility and Screen Recording grant status.
    • Adds permissions grant command to request macOS Accessibility and Screen Recording grants for Cua Driver.
    • Adds hermes computer-use status command to surface when cua-driver repair is required.
    • Supports HERMES_CUA_DRIVER_CMD environment variable to pin a custom cua-driver binary, which Hermes never replaces during auto-repair.

    +3 more show less

    • Supports --capability-manifest and --approve-capability-manifest flags passed to Cua in bounded mode.
    • cua-driver skills install now auto-detects Hermes and links Cua's skill pack into the Hermes skills directory.
    • Adds an interactive session picker with search, resume, lifecycle status tags (done, intr, err, empty), message counts, and delete-with-confirmation.

    8 details in Everything

14. v2026.8.16 Aug 16, 2026 · issue 048

    Hermes v2026.8.16 adds key_cmd credential sourcing, lean compaction mode, multi-source agent fan-out, and plugin SDK exports 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. and toolset config panels.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.8.16 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.8.16

    └──▷ USE IT
    Use a short-lived SSOSSOSingle sign-on, using one identity provider to sign in to multiple applications. It reduces separate passwords and centralizes account and access management. token for an enterprise gateway without storing a static key in .env — the command runs once per token lifetime and the fresh bearer is used on every request.
    yaml copy

    providers:
      my-gateway:
        base_url: https://gateway.internal.example.com/v1
        api_mode: chat_completions
        key_cmd: my-auth-cli print-token --profile prod

    • Adds key_cmd credential source to provider config, letting operators specify a shell command that prints a short-lived token instead of storing a static key — output can be a bare token or JSON with access_token and optional expires_in fields; token is cached until near expiry.
    • Exports McpTab, ToolsetConfigPanel, and HermesGateway type from @hermes/plugin-sdk, plus a host.getGateway() method returning the live $gateway instance, so desktop plugins can render full credential, OAuth, and MCP server configuration panels matching Settings → Capabilities.
    • Adds multi-source agent support to the desktop app — sockets, roster, SDK, and fan-out updates are now handled end-to-end.
    • Adds user-facing authorization flow for cua-driver browser attachment in computer-use mode.
    • Shows per-turn wall-clock duration in the desktop transcript.

    +1 more show less

    • Adds INFO log records for 'prompt accepted / turn finished' on the Desktop/TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. turn path.

    6 details in Everything

15. docs update Aug 14, 2026 · issue 045

    Hermes now routes ~/.ssh/config edits through an approval prompt instead of hard-blocking them.

    • Promotes ~/.ssh/config from hard-blocked to approval-gated: edits to the SSH client config (host aliases, ProxyJump, VS Code Remote-SSH targets) now go through the same approve-once/session/always prompt used for other ~/.ssh writes, rather than a flat refusal.
    • Non-interactive callers (ACP file bridge, background jobs with no human channel) still fail closed when attempting ~/.ssh/config writes.
    • Private keys, authorized_keys, and all other material under ~/.ssh remain hard-blocked regardless of approval state.

    3 details in Everything

16. docs update Aug 14, 2026 · issue 045

    Hermes installer now pre-installs cua-driver for Computer Use toolset; use --skip-computer-use to opt out.

17. docs update Aug 14, 2026 · issue 045

    Hermes gains a plugin CLI with install, search, doctor, and pack commands for managing community plugins.

    └──▷ TRY IT
    Pin a community plugin to an exact immutable commit SHA so your environment never drifts on upgrade.
    copy

    $ hermes install my-plugin --ref a3f1c2e4b5d6789012345678901234567890abcd

    Find plugins that expose a specific capability before committing to an install.
    copy

    $ hermes search --capability network_access --json

    Validate a plugin in CI and fail the build immediately if the manifest or loader registration has errors.
    copy

    $ hermes doctor ./my-plugin --ci

    • Adds install <identifier> [--force] [--ref COMMIT_SHA] command to install plugins from a Git URL or bare community index name, with --ref accepting a full 40-character commit SHA to pin an exact immutable revision.
    • Adds search [term] [--json] [--capability CAP] [--refresh] command for fuzzy searching the community plugin index by name, description, or tags, fetched from plugins.index_url and cached under ~/.hermes/cache/ for 24 hours with offline fallback.
    • Adds doctor [path-or-id] [--ci] command to validate a native plugin through the real manifest parser, loader, and registration path; --ci flag exits with code 1 on errors for CI pipeline integration.
    • Adds pack install <path-or-url> [--force] command to install a hermes-pack.yaml plugin pack — a declarative set of plugins each pinned to an exact 40-character commit SHA — with a mandatory review screen and per-plugin capability consent; exits non-zero on any plugin failure.
    • Adds pack export [--enabled-only] [--name NAME] command to emit a hermes-pack.yaml on stdout from the current install, recording repo and exact SHA for each git-installed plugin while stripping secrets, capability grants, and allow_* gates.

    +3 more show less

    • Adds pack show <path-or-url> command to dry-run parse, validate, and display a pack without installing anything.
    • Git installs record source, exact revision, and pin status in the plugins/.install-metadata.json sidecar file; config, environment values, secrets, and capability grants are never stored there.
    • Unpinned plugins can be updated with a pull command; pinned plugins require --force --ref <new-commit> to move to a new revision.

    8 details in Everything

18. docs update Aug 14, 2026 · issue 045

    Hermes adds a concise messaging verbosity option alongside all, result, error, and off modes.

19. v2026.8.13 Aug 13, 2026 · issue 045

    Hermes v2026.8.13 adds profiles.list/profiles.create ws RPCs, setup_mcp consent tool, image.generate RPCRPCRemote Procedure Call, a protocol that lets a program execute a function on another machine as if it were local, giving cyber tools a structured way to communicate across processes or hosts without managing raw sockets., and per-route webhook toolset overrides.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.8.13 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.8.13

    └──▷ TRY IT
    List all profiles from a desktop plugin to build a roster UI, including a last-session preview per profile.
    copy

    $ { "method": "profiles.list", "params": { "include_sessions": true } }

    • Adds profiles.list and profiles.create ws JSON-RPC methods (in tui_gateway/methods_profiles.py), giving desktop plugins the ability to enumerate profiles (name, path, model/provider, description, skill_count, optional last_session preview via include_sessions) and create new profiles without going through the dashboard REST router.
    • Adds setup_mcp tool that renders an inline 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 card in the desktop transcript using a clarify-style blocking bridge, with composer draft suggestions surfaced as brand pills.
    • Adds image.generate ws RPC for plugin surfaces.
    • Adds per-route toolset overrides for webhook agent runs.
    • Surfaces model drift impact in the Desktop cron view.

    +4 more show less

    • Adds cron and blueprint recipes to the sidebar nav rail.
    • Labels the agent's recommended choice on clarify surfaces across the desktop UI.
    • Adds skill-match, connection-repair, and recurrence-to-cron suggestion providers to the desktop composer.
    • Suppresses composer suggestions the user has repeatedly ignored.

    9 details in Everything

20. docs update Aug 11, 2026 · issue 042

    Hermes Agent docs reveal CLI surface: install flags, config commands, and gateway setup for multi-platform deployments.

    └──▷ TRY IT
    Set up Hermes with the Nous Portal in one step to get 300+ models and the Tool Gateway without per-tool API key configuration.
    copy

    $ hermes setup --portal

    Install Hermes as an unprivileged service user on a headless server where browser automation is not needed.
    copy

    $ curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash -s -- --skip-browser

    • Adds hermes setup --portal to log in to the Nous Portal, set Nous as the 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, and enable the Tool Gateway (web search, image generation, TTS, cloud browser) in one command.
    • Adds --skip-browser flag to the install script (install.sh) to bypass Playwright/Chromium installation for headless deployments that don't need browser automation.
    • Adds hermes model command to interactively choose an LLM provider and model, including custom endpoints for local or self-hosted models.
    • Adds hermes tools command to configure which tools are enabled.
    • Adds hermes gateway setup command to connect messaging platforms (Telegram, Discord, Slack, and others).

    +9 more show less

    • Adds hermes config set and hermes config get commands to read and write individual configuration values (e.g. hermes config set OPENROUTER_API_KEY your_key).
    • Adds hermes config check and hermes config migrate commands for post-update configuration validation and migration.
    • Adds hermes doctor diagnostic command that surfaces missing dependencies, the detected install method (git installer, Docker, or NixOS), and exact remediation steps.
    • Adds hermes desktop command to launch the desktop application after a CLI-only install.
    • Adds hermes setup full setup wizard to configure all options at once.
    • Supports root-mode FHS install layout (/usr/local/lib/hermes-agent/, /usr/local/bin/hermes) for shared-machine and system-wide deployments, with per-user config still stored under ~/.hermes/ or $HERMES_HOME.
    • Supports non-sudo / system service user installs with graceful Playwright degradation: Chromium is installed into the user's local cache and the admin command needed for system libraries is printed.
    • Auto-detects install method (git installer, Docker, or NixOS) and surfaces the matching update command via hermes update.
    • Tool Gateway provides web search, image generation, TTS, and cloud browser access under a single Nous Portal subscription covering 300+ models.

    14 details in Everything

21. docs update Aug 11, 2026 · issue 042

    Hermes adds hermes import-agent claude-code migration, Cursor-style context variables, and per-user identity headers for remote 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.

    └──▷ TRY IT
    Migrate an existing Claude Code MCP setup — including servers, skills, and instructions — into Hermes in one step.
    copy

    $ hermes import-agent claude-code

    Send a per-user identity header to a multi-tenant remote MCP server so it can apply per-user rate limits or audit trails.
    yaml copy

    team_api: "https://mcp.team.example.com/mcp"
    identity_header:
      name: "X-User-Id"
      value_from: "profile"

    • Adds hermes import-agent claude-code command to migrate mcpServers blocks from ~/.claude.json — along with skills and instructions — into Hermes automatically.
    • Supports Cursor-style context variable substitution in MCP config: ${userHome}, ${workspaceFolder}, ${workspaceFolderBasename}, ${pathSeparator}, and ${/} (case-sensitive).
    • Adds identity_header config key for remote HTTP/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. MCP servers, supporting value_from: static (sends a literal value) or value_from: profile (sends the active Hermes profile name) — enabling per-user rate limits, audit trails, and multi-tenant routing. Ignored with a warning on stdio transports.

    3 details in Everything

22. docs update Aug 11, 2026 · issue 042

    Hermes adds per-session container isolation, gateway agent cache with memory budgeting, and client-side STT silence trimming

    ▶ ⤢
    └──▷ USE IT
    Isolate every conversation in its own throwaway container so no filesystem state leaks between gateway sessions.
    yaml copy

    container_persistent: false

    Cap gateway memory consumption on a busy multi-platform deployment by setting an explicit anonymous-RSS budget and protecting only the 10 most-active sessions from eviction.
    yaml copy

    agent_cache:
      max_size: 128
      idle_ttl_secs: 3600
      memory_high_mb: 4096
      max_evictions_per_pass: 16
      protect_recent: 10

    Disable client-side silence trimming when using a cloud STT provider to transcribe music or ambient audio where pauses are meaningful.
    yaml copy

    stt:
      cloud_trim_silence: false

    • Adds container_persistent: false on the Docker backend to give every chat session its own fresh container, destroyed when the session closes or goes idle — no filesystem state, mounts, or background processes carry over between sessions.
    • Adds gateway_turn_lease_timeout (default 1800 seconds) to cap how long a turn waits for a session lease before the gateway fails closed and returns a rejection notice instead of queueing the message.
    • Adds agent_cache config block with max_size, idle_ttl_secs, memory_high_mb, max_evictions_per_pass, and protect_recent keys to bound the gateway's per-session agent cache by count, time, and anonymous RSS — evicting least-recently-used transcripts under memory pressure.
    • Adds memory_high_mb: auto to derive the gateway agent cache memory budget automatically from the container or systemd unit limit (MemoryMax / MemoryHigh), keeping it in sync without a second number.
    • Adds runtime.nofile_soft_limit config key to set the RLIMIT_NOFILE soft limit at server startup, clamped to the OS hard limit; set to 0 to disable the adjustment.

    +2 more show less

    • Adds stt.local.unload_after_idle_seconds to automatically release a local STT model from memory after idle time (e.g. 300 for 5 minutes), freeing GPUGPUA graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. memory on CUDACUDANVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. hosts.
    • Adds pre-upload silence trimming via ffmpeg for cloud STT providers (groq, openai, mistral, xai, elevenlabs, deepinfra) by default, controlled by stt.cloud_trim_silence (set to false to disable) and tunable with cloud_trim_threshold_db (default -40) and cloud_trim_keep_ms.

    7 details in Everything

23. docs update Aug 11, 2026 · issue 042

    Hermes adds session resume, kanban review routing, storage optimization, and database repair commands to its CLI.

    └──▷ TRY IT
    Resume the most recent session in a specific project directory without leaving your current shell location.
    copy

    $ hermes --in /path/to/project --resume latest

    Route a completed task to a specific reviewer for handoff without blocking it.
    copy

    $ hermes request-review TASK-42 --reviewer alice

    Reclaim disk space on a large Hermes database without risking session data loss.
    copy

    $ hermes optimize-storage

    • Adds --resume latest keyword to resume the most recent session, scoped to the current workspace.
    • Adds --in <dir> flag to change into <dir> before starting or resuming, scoping --resume latest and -c lookups to that directory's workspace.
    • Adds request-review <id> subcommand to move a task to review status with a reviewer handoff; supports --reviewer flag to reassign before dispatch.
    • Adds request-changes <id> <reason> subcommand for reviewer verdict: closes the review attempt and routes the task back to its original implementer.
    • Adds reopen-review <id>... subcommand to send review task(s) back for changes (review → ready/todo), with --reason flag to append a comment.

    +6 more show less

    • Adds optimize subcommand to reclaim disk space by merging FTS5FTS5A full-text search extension built into SQLite that lets tools index and query large bodies of text efficiently, with ranking and phrase-match support, without an external search engine. index segments and running VACUUM — non-destructive, no session data changes.
    • Adds optimize-storage subcommand to migrate the full-text search index to the compact v23 external-content layout, reclaiming a large fraction of state.db on large databases.
    • Adds repair subcommand to fix a malformed state.db schema (e.g. table messages_fts already exists) so hidden sessions reappear, with automatic backup.
    • Adds repair-routing subcommand to re-attach gateway conversations stranded after a restart; dry-run by default, with --max-gap-seconds N to tune the contiguity window.
    • Adds recover subcommand for offline, non-destructive recovery of a damaged state.db into a separate clean database.
    • Adds retitle-skills subcommand to regenerate titles for sessions opened with a /skill using what the user actually typed; lists changes unless a confirm flag is passed.

    11 details in Everything

24. docs update Aug 11, 2026 · issue 042

    Hermes now supports repo-local reviewer checklists loaded from .agents/checks/*.md before PR reviews.

    • Adds support for .agents/checks/*.md — per-repository, plain-markdown review checklists stored under .agents/checks/ (e.g. security.md, migrations.md, public-api.md) that Hermes loads before reviewing a change touching the matching area, giving review agents project-specific criteria instead of relying on generic prompts.

    1 detail in Everything

25. docs update Aug 11, 2026 · issue 042

    Hermes adds hermes plugins subcommands for installing, enabling, and managing native and portable Agent Plugins.

    └──▷ TRY IT
    Audit all installed plugins and remove one that is no longer needed.
    copy

    $ hermes plugins list
    hermes plugins remove plugin-name

    • Adds hermes plugins install owner/repository --no-enable to install a native or portable Agent Plugin package without immediately activating it.
    • Adds hermes plugins list to view all installed plugins.
    • Adds hermes plugins enable <plugin-name> and hermes plugins disable <plugin-name> to toggle individual plugins.
    • Adds hermes plugins update <plugin-name> and hermes plugins remove <plugin-name> for lifecycle management of installed plugins.
    • Portable Agent Plugins v1 packages load Agent Skills and stdio 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. entries and remain disabled until explicitly enabled via the opt-in workflow.

    +1 more show less

    • Session titles now appear as a gold badge pinned to the far-right edge of the UI, truncating long titles before displacing model and context fields.

    6 details in Everything

26. docs update Aug 11, 2026 · issue 042

    Hermes adds Actual Computer support for private inference and a hermes backup migration command.

    └──▷ TRY IT
    Run Hermes against a local Actual Computer daemon for fully air-gapped private inference.
    copy

    $ ACTUAL_BASE_URL=http://127.0.0.1:8080 hermes

    Move your Hermes configuration to a new machine without manually recreating profiles.
    copy

    $ hermes backup

    • Supports Actual Computer as a private inference cluster backend, configurable via ACTUAL_API_KEY (hosted relay) or ACTUAL_BASE_URL=http://127.0.0.1:8080 (local daemon, no key required).
    • Adds hermes backup command to migrate an entire setup or a single profile to another machine without rebuilding from scratch.

    2 details in Everything

27. docs update Aug 4, 2026 · issue 035

    Hermes Agent docs debut: installer, CLI commands, and Tool Gateway surface named for immediate use.

    └──▷ TRY IT
    Fastest onboarding: authenticate with Nous Portal, set the provider, and enable the full Tool Gateway (search, image gen, TTS, cloud browser) in a single step.
    copy

    $ hermes setup --portal

    Install Hermes on a headless server or unprivileged service account where browser automation is not needed, skipping the Playwright/Chromium download entirely.
    copy

    $ curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash -s -- --skip-browser

    Diagnose a broken install — surfaces missing dependencies, detected install method, and exact remediation steps.
    copy

    $ hermes doctor

    • Adds hermes setup --portal to log in, set Nous as the 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, and enable the Tool Gateway (web search, image generation, TTS, cloud browser) in one command.
    • Adds hermes model command to interactively choose an LLM provider and model, including custom endpoints for local/self-hosted models.
    • Adds hermes gateway setup command to connect messaging platforms (Telegram, Discord, Slack, and others).
    • Adds hermes tools command to configure which tools are enabled.
    • Adds hermes config set and hermes config get commands to read and write individual config values (e.g. hermes config set OPENROUTER_API_KEY your_key).

    +11 more show less

    • Adds hermes config check and hermes config migrate commands for inspecting and migrating configuration after updates.
    • Adds hermes doctor command to surface environment diagnostics, detected install method, and actionable fixes.
    • Adds hermes desktop command to launch the desktop app after a CLI-only install.
    • Supports --skip-browser flag on the install script to bypass Playwright/Chromium installation for headless deployments (curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash -s -- --skip-browser).
    • Supports root-mode FHS install layout (/usr/local/lib/hermes-agent/, /usr/local/bin/hermes) via sudo curl … | sudo bash for shared-machine deployments; per-user config still resolves under ~/.hermes/ or $HERMES_HOME.
    • Supports per-user install layout with code at ~/.hermes/hermes-agent/ and binary symlink at ~/.local/bin/hermes.
    • Supports non-sudo/system service user installs with graceful degradation: installer detects missing sudo, skips --with-deps, installs Chromium into the user's local Playwright cache, and prints the admin command needed for system libraries.
    • Auto-detects install method (git installer, Docker, or NixOS) and surfaces the matching update command via hermes update and environment summary via hermes doctor.
    • Installer automatically provisions all dependencies — Python 3.11 (via uv), Node.js v22, ripgrep, and ffmpeg — with no manual prerequisite setup required beyond git (and curl/xz-utils on Linux).
    • Nix/NixOS support available as best-effort via a Nix flakeNix flakeA Nix packaging construct that pins all dependencies in a lockfile and exposes a standardized interface, giving cyber tools fully reproducible builds across any machine running the Nix package manager., declarative NixOS module, and optional container mode (see nix-setup.md).
    • Android (Termux) supported as an install target via the standard curl installer.

    16 details in Everything

28. docs update Aug 4, 2026 · issue 035

    Hermes adds per-session stall watchdog, compression timeout caps, font customization, and auxiliary concurrency limits.

    ▶ ⤢
    └──▷ USE IT
    Prevent a slow or hung summary model from stalling a session indefinitely by setting both a streaming ceiling and an inactivity cutoff.
    yaml copy

    context_timeout_seconds: 30
    context_total_ceiling_seconds: 90

    Enable the session stall watchdog so users in busy multi-session deployments are notified when the agent appears wedged, without killing the turn.
    yaml copy

    agent:
      session_stall_timeout: 120  # seconds; 0 disables

    Cap auxiliary 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. concurrency to avoid amplifying retries during a provider incident across many simultaneous sessions.
    yaml copy

    # under your compression / title-generation auxiliary config
    max_concurrency: 2

    • Adds terminal.font_family config key to set the embedded terminal font (accepts a locally installed family name or CSS font stack, with JetBrains Mono as fallback); also editable at Settings → Appearance → Terminal Font.
    • Adds context_timeout_seconds config key to cap the inactivity budget for in-agent compress_context operations (conversation loop, preflight compaction, and manual compress); set to 0 to disable.
    • Adds context_total_ceiling_seconds config key to set an absolute pre-commit ceiling on in-agent compression even while summary tokens are still streaming; clamped to at least context_timeout_seconds.
    • Adds agent.session_stall_timeout config key (default 0 = disabled) for a notify-only gateway watchdog that logs a WARNING and sends a one-shot user notification when a session with a pending inbound follow-up has been idle beyond the threshold.
    • Adds max_concurrency cap for auxiliary compression LLM calls and title-generation calls to prevent retry bursts from overwhelming rate-limited providers across multi-session deployments (Discord/Telegram channels, multiple terminals).

    +1 more show less

    • Adds latency as an opt-in status-bar field (wall-clock turn duration, e.g. 22s / 1m05s) alongside the existing default set of ["model", "context_pct", "cwd"].

    6 details in Everything

29. docs update Aug 4, 2026 · issue 035

    Hermes CLI gains prompt stashing with Ctrl+S and zero-cost shell mode via !<command>

    • Adds Ctrl+S prompt stash: parks the current composer draft and clears the input so you can send something else first; pressing Ctrl+S again on an empty composer restores the draft (with cursor position and attached images). Repeated presses build a stack — with two or more stashed, Ctrl+S opens a browse panel (↑/↓ to navigate, Enter to restore, D to discard, Esc/Ctrl+S to close). A 📌 N badge in the status bar shows the number of parked drafts.
    • Adds !<command> shell mode in the CLI: prefixing a line with ! (e.g. !git status, !pytest -x tests/cli) runs it as a shell command without invoking the model — no API call, no tokens, no latency, and no history entry. Dangerous commands (e.g. rm -rf) still go through the approval prompt. Non-zero exits print ! exited <code>. Shell mode is CLI-only; Gateway platforms (Discord, Telegram, Slack) and cron runs ignore it.
    • Adds /personality none (also accepts default or neutral) to remove any active personality overlay and return to the default.

    3 details in Everything

30. v2026.8.3 Aug 3, 2026 · issue 035

    Hermes v2026.8.3 adds conversational voice with barge-in, 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. v1.0, signed outbound webhooks, grounded citations, a plugin SDK, and a wave of CLI power commands.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.8.3 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.8.3

    └──▷ TRY IT
    Instantly run a shell command from inside a Hermes CLI session without consuming a model turn.
    copy

    $ !git status

    Bootstrap a new project by having Hermes scan it and generate an AGENTS.md with its conventions and structure.
    copy

    $ /init

    Inspect what is consuming your context windowcontext windowThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. before a long autonomous run to decide whether to compress.
    copy

    $ /context

    • Adds !command shell escape to run a shell command instantly from the CLI without spending a model turn.
    • Adds /init CLI command to scan a project and generate or update an AGENTS.md file.
    • Adds /diff CLI command to show staged, all, or session changes from any surface.
    • Adds /context CLI command to break down exactly what is filling the context window.
    • Adds /focus CLI command for a reduced-output view with hidden-line recovery.

    +23 more show less

    • Adds hermes import-agent command to migrate a Claude Code or Codex CLI setup into Hermes in one command.
    • Adds hermes approvals suggest command to mine approval history into allowlist proposals.
    • Adds a new hermes tools category for STT configuration with GUIGUIA graphical user interface, controls such as windows, buttons and menus. It lets people use software visually rather than entering commands. toggles and dashboard dropdowns.
    • Raises the default 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. iteration limit from 90 to 500, enabling long autonomous runs that previously hit an artificial ceiling.
    • Adds streaming conversational voice mode with clause-by-clause TTS, barge-in interruption, and busy-aware silence detection across the CLI, desktop, and gateway adapters.
    • Adds on-device open-vocabulary wake-word detection with multi-profile voice routing and a stop keyword to end voice chat on any surface — no audio leaves the device while waiting.
    • Adds voice note support on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, and Weixin, with auto-TTS replies delivered in the platform-appropriate codec (opus with captions).
    • Adds OpenAI gpt-transcribe as a supported STT provider with unified language resolution.
    • Adds the grounded-citations skill, which matches quotes against actual page text, links citations to exact evidence, and includes a fact-checking mode for any document or claim.
    • Adds signed outbound webhooks that push HMACHMACA keyed hash that proves a message came from someone holding a shared secret and was not altered. Webhooks and API request signing commonly use it.-verified lifecycle events (session activity, turn completions, tool events) to any registered HTTP endpoint — no polling required.
    • Adds a bundled Agent-to-AgentAgent-to-AgentAgent2Agent, 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. (A2A v1.0) plugin so Hermes can discover, communicate with, and be driven by other A2A-compatible agents.
    • Adds desktop artifacts: versioned cards with sandboxed live preview in a right-rail viewer so generated HTML/apps run safely beside the chat.
    • Adds a plugin SDK with ctx.download for handing users files, floating pane placement, and multiple GUI windows; includes Kanban as the founding plugin.
    • Adds a global-hotkey quick-entry window to capture a thought into any session from anywhere in the OS.
    • Adds mid-turn redirect capability: typing a correction while the agent is working redirects the active turn, preserving work in flight and the original prompt.
    • Adds a compression-threshold configuration that is now settable per-model and in absolute tokens, with a guaranteed N-user-message tail so recent conversation always survives pruning.
    • Adds a consecutive-denial circuit breaker to stop misbehaving approval loops, plus a new approval gate for docker/podman daemon-redirect commands.
    • Adds an approval gate specifically for docker/podman daemon-redirect commands.
    • Adds proactive tool-result pruning, per-turn micro-compaction, and progress-aware timeouts to context compression for large-window models.
    • Adds truncated terminal output spill-to-file so the agent can read back output that exceeded the display limit.
    • Adds patch self-diagnosis for already-applied edits and whitespace mismatches, write_file on-disk content verification, and near-miss probing for searches that match nothing.
    • Adds prompt caching coverage for tool schemas on native Anthropic without history loss.
    • Reduces hermes -w cold start from ~14 s to ~1.8 s and makes hermes update no-ops 2–6 s faster.

    28 details in Everything

31. docs update Jul 31, 2026 · issue 031

    Hermes voice mode gains full-duplex barge-in: interrupt the agent mid-generation or mid-playback with speech or a stop phrase.

    └──▷ USE IT
    Tune barge-in sensitivity and grace period in config.yaml to reduce false triggers in a noisy environment.
    yaml copy

    voice:
      barge_in_threshold_multiplier: 2.5
      barge_in_grace_seconds: 0.8

    Stream live VAD diagnostics to stderr while running Hermes to dial in the right threshold for your microphone and room.
    copy

    $ HERMES_VOICE_DEBUG=1 hermes

    • Adds full-duplex barge-in: microphone stays live throughout the agent's turn, letting you interrupt mid-generation (before audio plays) or mid-playback by speaking.
    • Adds stop phrase support: saying stop during generation or playback both cuts the current turn and ends the voice chat session.
    • New voice.barge_in_threshold_multiplier config key scales the speech-detection trigger relative to a quiet-room noise floor calibrated at turn start.
    • New voice.barge_in_grace_seconds config key (default 0.5) suppresses false trips immediately after playback begins.
    • New HERMES_VOICE_DEBUG=1 environment variable streams per-block VAD diagnostics (calibrated floor, RMS, trip decisions) to stderr for live tuning.

    5 details in Everything

32. docs update Jul 31, 2026 · issue 031

    Hermes 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. elicitation support, routing structured mid-tool-call input requests to CLI, Telegram, and Slack approval surfaces.

    └──▷ USE IT
    Enable elicitation for a specific MCP server and tune the response timeout so users on async surfaces (e.g. Slack) have enough time to answer before the server gives up.
    yaml copy

    elicitation:
      # default: true
      enabled: true
      timeout: 300  # seconds to wait for your answer (default: 300)

    • Supports MCP elicitation/create protocol (requires mcp Python SDK ≥ 1.11.0), allowing MCP servers to request structured user input mid-tool-call.
    • Routes form-mode elicitations through the existing approval surface — interactive CLI/TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. prompt or approval buttons on Telegram and Slack gateways.
    • Adds per-server elicitation configuration with an elicitation key, including an enable/disable toggle and a configurable timeout (default 300 seconds).
    • Tracks per-server elicitation metrics (requests, accepted, declined, errors) on the handler.
    • MCP server reads conversation data directly from ~/.hermes/state.db as the primary session store, with sessions.json as a legacy fallback.

    5 details in Everything

33. docs update Jul 31, 2026 · issue 031

    Hermes gains a journey command to visualize, prune, and edit the full memory/skill timeline in CLI, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., and desktop.

    └──▷ TRY IT
    Replay how Hermes built its knowledge over time to audit what it has learned before a sensitive engagement.
    copy

    $ hermes journey --play --fps 10

    Find and delete a stale or incorrect memory chunk to stop Hermes from using bad context in future sessions.
    copy

    $ hermes journey list
    hermes journey delete memory:conversation:42 -y

    Correct a miscaptured skill directly in your editor so Hermes applies the right procedure going forward.
    copy

    $ hermes journey edit "port-scan-workflow"

    • New hermes journey command renders a timeline of all learned skills and memory entries in the terminal, with aliases hermes learning and hermes memory-graph.
    • Animatable build-up playback via --play flag (with --fps, --width, --height, --no-color, --json controls) lets practitioners replay exactly how Hermes's knowledge was assembled.
    • New hermes journey list subcommand lists all node IDs — skill names and memory:<source>:<index> identifiers — for inspection and scripting.
    • New hermes journey delete <node> subcommand removes nodes: skills are archived (restorable), memory chunks are permanently removed; -y skips confirmation.
    • New hermes journey edit <node> subcommand opens a skill or memory chunk in $EDITOR for direct correction.

    +1 more show less

    • TUI gains /journey overlay (aliases /learning, /memory-graph) and the desktop app gains an interactive Star Map panel — both exposing the same list/delete/edit operations.

    6 details in Everything

34. docs update Jul 31, 2026 · issue 031

    Hermes adds Vercel Sandbox as a cloud microVM execution backend with snapshot-backed filesystem persistence.

    • Adds vercel_sandbox as a new terminal backend option, enabling Hermes to run commands inside Vercel cloud microVMs alongside existing backends (docker, ssh, singularity, modal, daytona).
    • Supports snapshot-backed filesystem persistence across sandbox recreation, preserving Hermes-synced credentials, skills, and cache files for the same task.
    • Supports runtimes node24, node22, and python3.13 for the Vercel Sandbox backend, with /vercel/sandbox as the default remote workspace root.
    • Enables authentication via long-lived 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. (VERCEL_TOKEN, VERCEL_PROJECT_ID, VERCEL_TEAM_ID) for hosted deployments (Render, Railway, Docker, etc.) or short-lived 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. tokens (VERCEL_OIDC_TOKEN) for local development.

    4 details in Everything

35. docs update Jul 31, 2026 · issue 031

    Hermes adds iMessage/SimpleX support, per-channel model overrides, multi-select clarify prompts, and a tool-call audit log mode.

    └──▷ USE IT
    Run different models per channel from one gateway — e.g. a fast model in #daily and a frontier model with a specialist prompt in #dev.
    yaml copy

    # ~/.hermes/gateway-config.yaml
    channel_overrides:
      "123456789012345678":
        model: anthropic/claude-sonnet-4.6
        provider: anthropic
        system_prompt: "You are the #dev channel code-review specialist."
      "987654321098765432":
        model: openai/gpt-5-mini

    Search past sessions by title or ID when the session list has grown long.
    copy

    $ /sessions search auth-bypass

    Suppress tool-progress chat noise and capture a full tool-call audit trail to disk instead.
    yaml copy

    # ~/.hermes/gateway-config.yaml
    display:
      tool_progress: log

    • Adds Photon (iMessage) and SimpleX Chat as supported messaging platforms.
    • Introduces Hermes Relay, an experimental connector system that fronts Discord, Telegram, Slack, and WhatsApp through external connectors with per-handshake capability negotiation.
    • Adds /sessions search <query> to filter past sessions by title or session-id, and /sessions all for admin-level cross-origin session listing.
    • Persists /model overrides across gateway restarts — the model/provider choice is rehydrated from the session store on restart; adds --global flag to write the override to global config and --once to apply it for a single turn only.
    • Enables per-channel model and system prompt overrides via channel_overrides in ~/.hermes/gateway-config.yaml, allowing different models and personas per channel or thread from a single gateway.

    +4 more show less

    • Adds multi-select support to Clarify questions — agents can let users pick several options at once via comma-separated numbers on messaging platforms or Space/Enter checkboxes in the CLI/TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency..
    • Adds display.tool_progress: log mode that routes tool-call progress to a rotating audit file (~/.hermes/logs/tool_calls.log) instead of chat messages, with secret redaction.
    • Adds configurable status phrase catalogs via ~/.hermes/status_phrases.yaml or ~/.hermes/status_phrases/ directory, supporting append or replace modes.
    • Adds Microsoft GraphMicrosoft GraphMicrosoft's unified REST API for accessing data and services across Microsoft 365, Azure AD, and related cloud products, giving cyber tools a single authenticated endpoint to query users, emails, files, and sign-in logs. Webhook Listener and ntfy as supported integrations.

    9 details in Everything

36. docs update Jul 31, 2026 · issue 031

    Hermes adds Vercel Sandbox backend, in-place session compaction, and Verify-on-Stop coding enforcement

    └──▷ USE IT
    Run Hermes in a Vercel cloud microVM with a Python 3.13 runtime for isolated, snapshot-backed task execution.
    yaml copy

    # hermes config (hermes.yaml)
    terminal:
      backend: vercel_sandbox
      vercel_runtime: python3.13

    Enforce coding verification before Hermes accepts a final answer on any turn that edits code, with a cap of 2 nudges per turn.
    yaml copy

    # hermes.yaml
    verify_on_stop: true
    max_verify_nudges: 2

    Use 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. token from the Vercel CLI for one-off local development sessions without configuring long-lived credentials.
    copy

    $ VERCEL_OIDC_TOKEN="$(vc project token <project-name>)" hermes chat

    • Adds vercel_sandbox terminal backend: runs agent shell commands in Vercel cloud microVMs with snapshot-backed filesystem persistence across sessions.
    • Supports node24, node22, and python3.13 runtimes for the Vercel Sandbox backend via terminal.vercel_runtime.
    • Adds remote-to-host state sync on teardown for cloud backends (Modal, Daytona, Vercel Sandbox): credential files, skills, and cache are pushed into the sandbox and synced back to host paths on session end.
    • Adds in_place compaction mode: compacts session context without rotating the session ID, preserving a single durable conversation identity with soft-archived pre-compaction turns still searchable via session_search.
    • Adds Verify-on-Stop (verify_on_stop): refuses to accept a final answer on turns where code was edited but no verification evidence (test run, build, lint) was produced, with configurable nudge cap and auto surface-aware mode.

    +6 more show less

    • Adds verify_guidance and max_verify_nudges config keys to control the missing-evidence nudge message and cap on consecutive continue nudges per turn.
    • Adds coding_instructions config key for standing project-wide coding rules appended to the coding brief.
    • Raises the local provider stream stale timeout ceiling to 900 seconds (configurable via agent.local_stream_stale_timeout or HERMES_LOCAL_STREAM_STALE_TIMEOUT).
    • Supports Cursor-style ${env:VAR_NAME} SecretRef syntax, allowing 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. or provider config snippets copied from Cursor/Claude to work unchanged in the mcp_servers block.
    • Adds timestamp_format config key (strftime string) to control the format of per-message timestamps in the CLI/TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. transcript.
    • Adds agent.clarify_timeout as the canonical key for configuring how long the gateway waits for a user's clarifying-question response.

    11 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The config migration sets verify_on_stop to false on existing installs — treat off as the effective default and opt in explicitly if you want verification enforcement.
37. docs update Jul 31, 2026 · issue 031

    Hermes adds usage-cost reporting, new CLI commands, config inspection, relay enrollment, and uninstall tooling.

    └──▷ TRY IT
    Track per-run 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 in a CI pipeline — get a JSON cost breakdown even when the run fails.
    copy

    $ hermes --oneshot "summarize this repo" --usage-file /tmp/usage.json && jq .estimated_cost_usd /tmp/usage.json

    Inspect the current default model without opening a config file, useful for auditing or scripting environment checks.
    copy

    $ hermes config get model.default --json

    Preview what a full uninstall would remove before committing — safe to run on shared or production machines.
    copy

    $ hermes uninstall --full --dry-run

    • New --usage-file <path> flag writes a machine-readable JSON usage report (cost, token counts, API calls, session metadata) after any --oneshot run — even on failure — for pipeline spend tracking.
    • New hermes approvals command mines approval history into allowlist proposals.
    • New hermes skin command lists, switches, and tweaks display skins.
    • New hermes console command opens the safe Hermes command console.
    • New hermes journey command exposes a timeline of learned skills and memories over time.

    +6 more show less

    • New hermes config get <key> [--json] subcommand prints a single config value by dotted key with optional machine-readable output.
    • New hermes config unset <key> subcommand removes a config key and reverts it to the built-in default.
    • New hermes uninstall [--full] [--gui] [--dry-run] [--yes] command with --dry-run mode that prints what would be removed without changing anything.
    • Experimental relay enrollment available via ai-gateway to register a gateway with a relay connector and save credentials for connector-backed platforms.
    • Supports installing or checking the Computer Use (cua-driver) backend on macOS, Windows, and Linux.
    • New command to start the Hermes backend server in headless mode, powering the desktop app and remote backends.

    11 details in Everything

38. docs update Jul 31, 2026 · issue 031

    Hermes adds speech-to-text across eight providers with voice message transcription on Telegram, Discord, and WhatsApp.

    • Supports speech-to-text via eight providers: local faster-whisper (free, on-device), a local command wrapper, Groq, OpenAI Whisper API, Mistral, xAI, ElevenLabs Scribe, and DeepInfra.
    • Enables voice message transcription across Telegram, Discord, WhatsApp, and other messaging platforms.

    2 details in Everything

39. docs update Jul 31, 2026 · issue 031

    Hermes adds iMessage/SimpleX support, per-channel model overrides, multi-select clarify prompts, and a tool-call audit log mode.

    └──▷ TRY IT
    Search through a long session history to quickly resume a past investigation by keyword.
    copy

    $ /sessions search log4shell

    Run a cheap fast model in a general channel and a frontier model with a specialist prompt in a dev channel from a single gateway.
    yaml copy

    # ~/.hermes/gateway-config.yaml
    channel_overrides:
      "123456789012345678":
        model: openai/gpt-5-mini
      "987654321098765432":
        model: anthropic/claude-sonnet-4.6
        provider: anthropic
        system_prompt: "You are the #dev channel code-review specialist."

    Capture a full tool-call audit trail without any chat noise during an automated gateway session.
    yaml copy

    # ~/.hermes/gateway-config.yaml
    display:
      tool_progress: log

    • Adds Photon (iMessage) and SimpleX Chat as supported messaging platforms.
    • Introduces Hermes Relay (experimental), a connector system that fronts Discord, Telegram, Slack, and WhatsApp through external connectors with per-handshake capability negotiation.
    • Adds /sessions search <query> (alias find) to filter past sessions by title or session-id, and /sessions all for admin cross-origin session listing.
    • Persists /model overrides across gateway restarts — the model/provider choice is rehydrated from the session store on startup without writing credentials to disk.
    • Adds /model <name> --once to apply a model override for a single turn only, and /model <name> --global to write the override to global config.

    +5 more show less

    • Adds per-channel channel_overrides in ~/.hermes/gateway-config.yaml, allowing different models and system prompts per channel or thread from a single gateway.
    • Enables multi-select clarify questions: agents can prompt users to pick multiple options at once, rendered as comma/space-separated replies on messaging platforms and checkboxes in CLI/TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency..
    • Adds display.tool_progress: log mode that routes all tool-call progress to a rotating audit file (~/.hermes/logs/tool_calls.log) instead of chat messages, with secret redaction.
    • Supports configurable gateway status phrases via ~/.hermes/status_phrases.yaml or files under ~/.hermes/status_phrases/, with append or replace merge modes.
    • Adds Microsoft GraphMicrosoft GraphMicrosoft's unified REST API for accessing data and services across Microsoft 365, Azure AD, and related cloud products, giving cyber tools a single authenticated endpoint to query users, emails, files, and sign-in logs. Webhook Listener, LINE, and ntfy as supported integrations.

    10 details in Everything

40. docs update Jul 29, 2026 · issue 029

    Hermes Agent: terminal-native autonomous coding agent with persistent memory, 21+ messaging platforms, multi-backend execution, 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./ACP support.

    └──▷ TRY IT
    Install Hermes Agent on Linux, macOS, WSL2WSL2A Microsoft-maintained compatibility layer in Windows 10/11 that runs a real Linux kernel inside a lightweight VM, letting cyber tools built for Linux execute natively on Windows without dual-booting or a full VM., or Termux in a single step.
    copy

    $ curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

    • Supports 21+ messaging platforms natively, including Telegram, Discord, Slack, WhatsApp, Signal, Email (IMAP/SMTPSMTPThe application-layer protocol used to send email between servers and from clients to servers; cyber tools use it to relay alerts, phishing simulations, or exfiltrated data over standard mail infrastructure.), SMSSMSShort Message Service, the protocol built into cellular networks for sending text messages between phones; cyber tools use it for alerting, two-factor authentication delivery, or phishing simulation over mobile channels. via Twilio, Matrix, Mattermost, Home Assistant, and IRC/Microsoft Teams via plugins.
    • Runs on local, Docker, SSH, Daytona, Modal, or Singularity execution backends.
    • Connects to Nous Portal, 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., OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint as an 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.
    • Persistent memory across sessions via MEMORY.md, USER.md, and session search; external memory provider plugins include Honcho, Mem0, Hindsight, and others.
    • Agent-created Skills System with background Curator maintenance (usage tracking, staleness detection, LLM-driven review).

    +19 more show less

    • Exposes an OpenAI-compatible API server for use with any frontend.
    • Supports MCP (Model Context ProtocolModel Context ProtocolModel 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.) for connecting to external tool servers, with per-tool load control.
    • Supports ACP (Agent Context Protocol) for use inside compatible editors and collaboration platforms.
    • Kanban multi-agent coordination backed by SQLite for durable task board management across multiple Hermes profiles.
    • Task delegation via delegate_task spawns isolated child agents for parallel workstreams.
    • Cron job scheduling with natural-language task definitions and per-job skill attachment.
    • Persistent Goals feature keeps the agent working across turns until a standing goal is complete.
    • Checkpoint and rollback system uses shadow git repos and automatic snapshots as filesystem safety nets.
    • Git worktree support enables multiple Hermes agents to operate safely on the same repository in isolation.
    • Real-time voice mode available in CLI, Telegram, and Discord (DMs, text channels, and voice channels).
    • Inline @-syntax for attaching files, folders, git diffs, and URLs directly into messages as context references.
    • Webhook ingestion from GitHub, GitLab, and other services to trigger agent runs.
    • Batch processing for generating agent trajectories at scale with parallel execution, checkpointing, and toolset distributions.
    • Programmatic Python code execution with RPCRPCRemote Procedure Call, a protocol that lets a program execute a function on another machine as if it were local, giving cyber tools a structured way to communicate across processes or hosts without managing raw sockets. tool access to collapse multi-step workflows into a single turn.
    • Plugin system for extending Hermes with custom tools, hooks, and integrations.
    • Honcho memory integration for dialectic reasoning, multi-agent user modeling, and deep personalization.
    • Provider routing, fallback providers, and credential pools for resilient multi-provider LLM setups.
    • Runs on Android via Termux and supports NixOS declarative deployment via a Nix module.
    • Embeddable as a Python library (AIAgent) for use in scripts, web apps, or automation pipelines without the CLI.

    24 details in Everything

41. docs update Jul 29, 2026 · issue 029

    Hermes adds approvals suggest to mine past approval decisions into allowlist proposals automatically.

    └──▷ TRY IT
    Preview which frequently-approved commands qualify for your allowlist before committing any changes.
    copy

    $ hermes approvals suggest --days 60 --min-count 3

    Apply specific proposals (e.g. items 1 and 3) from the ranked list into your command_allowlist after reviewing the dry-run output.
    copy

    $ hermes approvals suggest --apply 1,3

    Pipe proposal output into a script or audit tool for automated allowlist review in CI.
    copy

    $ hermes approvals suggest --json --days 90 --limit 20

    • New hermes approvals suggest command mines the session database for dangerous-classified commands you previously approved and proposes them as command_allowlist additions, ranked by approval frequency.
    • Supports --apply N[,M...] flag to selectively write chosen proposals into the allowlist, with destructive command classes permanently excluded regardless of approval count.
    • Adds --days N, --min-count N, --limit N, and --db PATH flags to control the history window, minimum approval threshold, result count, and database path for proposal generation.
    • Supports --json flag for machine-readable output of allowlist proposals.
    • Extends dangerous-command detection to cover Docker and Podman remote/daemon-redirect patterns, including docker -H/--host/--context, DOCKER_HOST=, DOCKER_CONTEXT=, docker context use, and podman --remote/-r/--url/--connection/--identity/CONTAINER_HOST=.

    5 details in Everything

42. docs update Jul 29, 2026 · issue 029

    Hermes adds voice mode support, enabled via hermes setup voice.

    └──▷ TRY IT
    Enable voice mode so you can interact with the Hermes agent using speech.
    copy

    $ hermes setup voice

43. docs update Jul 29, 2026 · issue 029

    Hermes voice mode gains barge-in interruption, stop-phrase ending, real-time streaming TTS, and per-call voice-design instructions.

    └──▷ USE IT
    Customize which spoken or typed phrases end a voice session, and add a second farewell phrase, so users can say goodbye hermes to exit hands-free.
    yaml copy

    voice:
      stop_phrases: ["stop", "goodbye hermes"]

    Disable barge-in when recording in a noisy environment where accidental interruptions would disrupt the agent's replies.
    yaml copy

    voice:
      barge_in: false

    Apply expressive voice-design (whispering, emotional tone) on a per-call basis when using gpt-4o-mini-tts via the text_to_speech tool.
    json copy

    # In a tool call to text_to_speech:
    {
      "text": "The operation completed successfully.",
      "instructions": "Speak in a calm, reassuring tone with a slight whisper."
    }

    • Adds voice stop-phrase detection: saying or typing exactly a configured phrase (default: stop) ends the voice chat hands-free; customizable via voice.stop_phrases or set to [] to disable.
    • Adds barge-in support: a voice-activity monitor cuts agent playback the moment the user starts talking and returns to recording immediately, with automatic noise-floor calibration to avoid speaker bleed self-triggering; disable with voice.barge_in: false.
    • Adds agent interruption awareness: when barge-in cuts a reply, the next message carries a note informing the model its spoken reply was cut off so it can respond naturally.
    • Adds real-time sentence-level TTS streaming: audio plays per sentence as the model generates text, with chunked PCM streaming (ElevenLabs, OpenAI) for lowest time-to-first-word across CLI, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., and desktop.
    • Adds optional per-call instructions argument to the text_to_speech tool for voice-design (tone, emotion, pacing, accent, whispering) with gpt-4o-mini-tts and OpenAI-compatible voice-design servers via tts.openai.base_url.

    +2 more show less

    • Adds auto-end of voice chat after three consecutive silent cycles with no speech detected.
    • Adds optional 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.-639-1 language hint for STT via the language config field; falls back to HERMES_LOCAL_STT_LANGUAGE or auto-detect.

    7 details in Everything

44. docs update Jul 29, 2026 · issue 029

    Hermes adds Figma remote 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. support and fnmatch glob patterns for tool allow/exclude lists.

    └──▷ TRY IT
    Install and authenticate the Figma remote MCP server without any special OAuth tricks.
    copy

    $ hermes mcp install figma && hermes mcp login figma

    Exclude entire Cloudflare product areas by glob pattern when a server exposes thousands of tools.
    copy

    $ cloudflare "https://mcp.cloudflare.com/mcp?codemode=false"
      "*_radar_*"
      "*_accounts_dlp_*"
      "*_zones_web3_*"

    • Adds native Figma remote MCP integration: Hermes auto-sets oauth.client_name: "Claude Code" for mcp.figma.com so OAuth registration succeeds without manual workarounds.
    • Supports fnmatch-style glob patterns in MCP tool allow/exclude lists, enabling bulk filtering across large tool surfaces (e.g. Cloudflare's ~3,300-tool API MCP) without listing each endpoint individually.

    2 details in Everything

45. docs update Jul 29, 2026 · issue 029

    Hermes adds mid-turn agent redirection via Buzz, letting you correct an active turn without losing completed work.

    • Adds Buzz messaging mode: send a message while the agent is working to redirect its active turn, preserving reasoning, partial text, and prior tool results as a checkpoint.
    • Running tools complete safely when a Buzz redirect is issued — the correction is applied at the next tool-result boundary rather than killing the tool mid-execution.
    • Supports configurable busy acknowledgment suppression so the confirmation message can be hidden without changing input-handling behavior.

    3 details in Everything

46. docs update Jul 29, 2026 · issue 029

    Hermes adds idle compaction, proactive tool-result pruning, per-turn loop caps, focus view, and richer CLI turn summaries

    └──▷ USE IT
    Prevent context bloat from old tool outputs on large-window models by enabling proactive pruning at 48k tokens, leaving small results and low-yield prunes alone.
    yaml copy

    proactive_prune_tokens: 48000
    proactive_prune_min_result_chars: 8000
    proactive_prune_min_reclaim_tokens: 4096

    Cap runaway agent turns in a single conversation turn to avoid unbounded web search or subagent delegation loops.
    yaml copy

    loop_caps:
      max_web_searches: 10
      max_subagents: 5

    Pre-compact a long-lived Telegram thread automatically after 30 minutes idle so stale context isn't re-read on every returning message.
    yaml copy

    idle_compact_after_seconds: 1800

    • Adds threshold_tokens config to set an absolute token cap for compression triggers, ensuring compression fires no later than a set token count regardless of model context-window size.
    • Adds idle_compact_after_seconds opt-in trigger to compact accumulated session history when a conversation resumes after a configurable idle period, before the first reply.
    • Adds proactive_prune_tokens to enable deterministic, no-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. pruning of old tool-result payloads on large-window models where size-based compaction rarely fires.
    • Adds proactive_prune_min_result_chars and proactive_prune_min_reclaim_tokens controls to tune when the proactive prune acts and whether it commits changes.
    • Adds progress_notices opt-in to surface routine compression lifecycle statuses (start, trigger, retry, complete) to chat platforms such as Telegram, Discord, and Slack.

    +7 more show less

    • Adds hygiene_timeout_seconds, hygiene_total_ceiling_seconds, and hygiene_failure_cooldown_seconds to control pre-agent compression pass timeouts, total wait ceiling, and per-session failure cooldowns.
    • Adds auxiliary.stream_only_base_urls config to force streaming mode for auxiliary tasks (title generation, compression, web extraction) on endpoints that reject non-streaming requests.
    • Adds loop_caps config block with max_web_searches and max_subagents to enforce per-turn hard ceilings on web search calls and subagent spawns.
    • Adds display.turn_summary CLI option to print a one-line post-turn accounting footer showing wall time, tool calls grouped by verb, and file-edit deltas.
    • Adds display.spinner_token_flow CLI option to display live cumulative output token count on the spinner during a turn.
    • Adds display.focus_view (/focus command) for a reduced-output display mode that suppresses tool play-by-play without affecting conversation history, prompt caching, or request payloads.
    • Expands STT language resolution with per-provider stt.<provider>.language overrides and HERMES_LOCAL_STT_LANGUAGE environment variable support across all STT providers.

    12 details in Everything

47. docs update Jul 29, 2026 · issue 029

    Hermes adds an outbound credential-injection firewall (hermes egress) and a Claude Code/Codex CLI importer (hermes import-agent).

    └──▷ TRY IT
    First-time egress proxy setup so sandboxes never hold real API keys — inject credentials at the network boundary instead.
    copy

    $ OPENROUTER_API_KEY=… hermes egress setup && hermes egress start

    Rotate proxy tokens after a suspected token leak, then restart the daemon to apply them.
    copy

    $ hermes egress setup --rotate-tokens && hermes egress restart

    • New hermes egress command family: a TLS-intercepting outbound firewall (iron-proxy) for remote terminal sandboxes that swaps opaque proxy tokens for real upstream API credentials at the network boundary, so sandboxes never hold real keys.
    • Supports Bitwarden Secrets Manager as a credential source via hermes egress setup --from-bitwarden, with --no-bitwarden to revert to env-based credentials.
    • Hot-reload of the egress ruleset without dropping connections via hermes egress reload, using the loopback management API.
    • Egress proxy status inspection with hermes egress status, including --show-tokens to print proxy tokens in full (redacted by default).
    • New hermes import-agent command imports Claude Code (~/.claude) or OpenAI Codex CLI (~/.codex) setups into Hermes, mapping CLAUDE.md instructions to memory entries, Bash permission rules to command_allowlist/approvals.deny, 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 to mcp_servers, and skill directories — with a mandatory preview step and no credential import.

    5 details in Everything

48. docs update Jul 29, 2026 · issue 029

    Hermes gains three integration paths into Buzz, Block's Nostr-based human+agent collaboration workspace.

    • Adds Buzz integration with three paths: Buzz Desktop as a managed ACP runtime, the buzz-acp relay bridge for server-side Hermes identity hosting, and a native gateway joining Buzz channels with full Hermes memory, skills, approvals, and cron support.

    1 detail in Everything

49. docs update Jul 29, 2026 · issue 029

    Hermes adds mid-turn agent redirection via Buzz — correct a running agent without losing completed work or killing active tools.

    • Supports sending a message while the agent is working to redirect its active turn, restarting model generation while retaining reasoning, partial text, and completed tool results.
    • Running tools finish safely when a mid-turn correction is issued — the correction is applied at the next tool-result boundary rather than terminating the tool.
    • Offers configurable busy-message modes, including suppressing the busy acknowledgment notification.

    3 details in Everything

50. docs update Jul 29, 2026 · issue 029

    Hermes adds hands-free Hey Hermes wake word to trigger voice sessions in CLI, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., and desktop app.

51. docs update Jul 29, 2026 · issue 029

    Hermes CLI gains mid-turn agent redirection and a visual context-usage breakdown command.

    └──▷ TRY IT
    See exactly how your context windowcontext windowThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. is being consumed — useful before a long agentic task to check available headroom.
    copy

    $ /context all

    Cancel the current agent turn and any foreground tool work immediately.
    copy

    $ /stop

    • New /context command displays a visual glyph block grid plus a per-category token table covering system prompt, tools, skills, memory, conversation, and free space.
    • New /context all subcommand adds per-skill and per-toolset token costs to the context breakdown.
    • Enables mid-turn redirection: sending a message while the agent is working restarts model generation with completed reasoning and tool work preserved, without cancelling in-flight tools.
    • New /stop command cancels the active turn and its foreground work.

    4 details in Everything

52. docs update Jul 29, 2026 · issue 029

    Hermes adds an egress credential-injection proxy for Docker sandboxes, keeping real API keys out of reach.

    └──▷ TRY IT
    Spin up the egress proxy before launching a Docker sandbox to ensure the container never sees your real API keys.
    copy

    $ hermes egress setup && hermes egress start

53. docs update Jul 21, 2026 · issue 021

    Hermes Agent launches with CLI install, multi-platform support, desktop app, and Tool Gateway integration.

    └──▷ TRY IT
    Fastest onboarding for a new machine: log in to Nous Portal, select a model from 300+, and enable the Tool Gateway in a single command.
    copy

    $ hermes setup --portal

    Install on a headless server or unprivileged service account where browser automation is not needed, skipping the Playwright/Chromium download entirely.
    copy

    $ curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash -s -- --skip-browser

    Diagnose a broken environment after an upgrade — surfaces missing dependencies, the detected install method, and exact remediation commands.
    copy

    $ hermes doctor

    • Adds hermes setup --portal one-command onboarding that logs in, sets Nous as the 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, and enables the Tool Gateway (web search, image generation, TTS, cloud browser).
    • Supports install across Linux, macOS, WSL2WSL2A Microsoft-maintained compatibility layer in Windows 10/11 that runs a real Linux kernel inside a lightweight VM, letting cyber tools built for Linux execute natively on Windows without dual-booting or a full VM., Windows (native PowerShell), Android (Termux), and NixOS with a single installer script that auto-provisions Python 3.11, Node.js v22, ripgrep, and ffmpeg.
    • Adds root-mode FHS layout (/usr/local/lib/hermes-agent/, /usr/local/bin/hermes) for shared-machine and system-service deployments.
    • Adds --skip-browser installer flag to bypass Playwright/Chromium installation for headless environments that don't require browser automation.
    • Adds hermes doctor diagnostic command that reports what's missing, surfaces the detected install method, and prints remediation steps.

    +6 more show less

    • Adds hermes model interactive provider/model selector supporting custom endpoints for local and self-hosted models.
    • Adds hermes tools command to configure which agent tools are enabled.
    • Adds hermes gateway setup command to connect messaging platforms (Telegram, Discord, Slack, and others).
    • Adds hermes config set / hermes config get for reading and writing individual config values, plus hermes config check and hermes config migrate for upgrade hygiene.
    • Adds hermes desktop command to launch the desktop application after a CLI-only install.
    • Adds hermes update with install-method auto-detection (pip, git installer, Homebrew, Nix) that prints the correct update command for the detected path.

    11 details in Everything

54. docs update Jul 21, 2026 · issue 021

    Hermes CLI gains a token command to rotate 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. with optional Bitwarden validation.

    └──▷ TRY IT
    Rotate your Hermes 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. in a CI pipeline without interactive prompts or Bitwarden validation.
    copy

    $ hermes token --no-verify

    • New token command rotates the access token, validating it against Bitwarden before storing; rejected tokens leave the existing token unchanged.
    • Supports --no-verify flag to skip the Bitwarden probe for non-interactive or automated use.

    2 details in Everything

55. v2026.7.20 Jul 20, 2026 · issue 021

    Hermes v2026.7.20 slashes first-token latency 80%, adds password-manager secret sources, live subagent transcripts, and terminal subscription management.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.7.20 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.7.20

    └──▷ TRY IT
    Export a session history as a Hugging Face-ready trace with secrets scrubbed — useful for building fine-tuning datasets from real agent conversations.
    copy

    $ hermes sessions export --redact

    Manage your Nous subscription tier without leaving the terminal — view your plan, preview upgrade cost, and apply changes.
    copy

    $ /subscription

    Watch a subagent work in real time by tailing its live transcript as soon as it launches.
    copy

    $ tail -f <transcript-file>

    • Cuts cold-start first-turn latency ~80% (4.3s → 0.9s) across CLI, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., gateway, desktop, and cron.
    • Streams reasoning model thinking live by default (display.show_reasoning on) — no more waiting at a spinner.
    • Adds /subscription and /topup commands to manage your Nous plan and billing directly from the terminal or TUI.
    • Adds /deny <reason> command to tell the agent why a command was refused so it can course-correct.
    • Introduces user-defined deny rules that block commands even under yolo mode.

    +15 more show less

    • Enables smart approvals by default — an 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. reviewer independently assesses flagged commands instead of prompting the user each time.
    • Adds Bitwarden and 1Password (op:// references) as pluggable SecretSource providers, so API keys no longer need to live in plaintext .env files.
    • Makes delegate_task subagent dispatches produce live tail-able transcript files per child agent.
    • Adds a durable delivery-obligation ledger in state.db so final responses are redelivered on next boot if the gateway crashed before platform delivery (covers Telegram, Discord, Slack, and others).
    • Supports profile-based message routing in a single multiplexed gateway — route guilds, channels, or threads to fully isolated profiles.
    • Adds hermes sessions export with Markdown, Quarto, HTML, prompt-only, and Hugging Face trace output formats, a --redact secret-scrubbing pass, and filters by age, workspace, and platform.
    • Adds Fireworks AI and DeepInfra as first-class model providers, plus Upstage Solar.
    • Adds new model catalog entries: GPT-5.6 (Sol/Terra/Luna + Pro variants), grok-4.5 (GA), moonshotai/kimi-k3, claude-fable-5/claude-sonnet-5, tencent/hy3 (GA), and LM Studio JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. local model loading.
    • Adds max and ultra reasoning-effort tiers with per-model config overrides and per-slot effort in MoA presets.
    • Adds a terminal execution backend picker (local/docker/ssh/modal/daytona/singularity) with per-backend health probes in the Desktop Capabilities tab.
    • Adds a declarative memory provider panel with full-config modal in the Desktop settings.
    • Adds a per-job model picker in the desktop cron create/edit dialog.
    • Adds per-session color override and project-inherited color/icon in the desktop session sidebar.
    • Delivers 14× less markdown-splitter CPUCPUThe main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. via incremental block lexing, virtualized review-pane diffs, and eliminated per-token sidebar re-renders in the desktop app.
    • Scopes Vertex credentials away from subprocess env, adds a shared credential-read guard for media/vision/image-gen local file reads, adds webhook body-size caps across all aiohttp servers, and redacts bot tokens in Telegram transport errors.

    20 details in Everything

56. v2026.7.7 Jul 8, 2026 · issue 014

    Hermes v2026.7.7 adds session bulk-archive, expanded prune filters, 1Password secrets, one-shot usage reports, and ~80% faster first-token latency.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.7.7 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.7.7

    └──▷ TRY IT
    Bulk-archive recent CI smoke-test sessions by source without deleting them, then confirm what was hidden.
    copy

    $ hermes sessions archive --source cron --started-after 2026-07-01

    Capture a JSON token/cost usage report after a one-shot automation run for billing or audit purposes.
    copy

    $ hermes -z "summarize /var/log/auth.log" --usage-file usage.json

    • Adds hermes sessions archive subcommand to bulk soft-hide sessions without deleting them.
    • Expands hermes sessions prune with a full filter surface: time bounds (--started-before/--started-after), model, provider, user/chat origin, git branch, token count, cost, tool calls, and more — filters now match all ages when any attribute filter is supplied, dropping the implicit 90-day cutoff.
    • Prune dry-run preview now reports match count plus the oldest and newest matching session's age span.
    • Adds pluggable SecretSource interface with a multi-source orchestrator, enabling secrets to be resolved from multiple backends in priority order.
    • Adds 1Password (op://) as a native secret source, letting Hermes resolve credentials directly from 1Password vaults.

    +2 more show less

    • Adds --usage-file flag to hermes -z (one-shot mode) to write a JSON usage report after each run.
    • Cuts first-turn time-to-first-tokentime-to-first-tokenTime to first token, the wait between submitting a request and receiving the first generated token. It includes delays such as queuing and processing the prompt. by ~80% across CLI, gateway, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., desktop, and cron by moving blocking Discord capability detection off the critical path and caching it to disk.

    7 details in Everything

57. v2026.7.1 Jul 1, 2026 · issue 014

    Hermes v2026.7.1 adds MoA as a first-class model, /learn, /journey, /prompt, background subagent fan-out, Vertex AI, and scale-to-zero gateway.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.7.1 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.7.1

    └──▷ TRY IT
    Run your hardest prompt through a named MoA ensemble to see each frontier model's reasoning before the aggregator synthesizes a final answer.
    copy

    $ /moa

    Teach Hermes a repeatable workflow from a directory you just walked it through, so it's available as a skill next time without manual authoring.
    copy

    $ /learn ./scripts/deploy-workflow

    Inspect and prune everything Hermes knows about you — edit or delete any memory or skill from the timeline.
    copy

    $ /journey

    • Adds Mixture-of-Agents (MoA) presets as selectable virtual models under a moa provider, available in every model picker across CLI, TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., desktop, and gateway.
    • New /moa command runs a single prompt through the default MoA preset as a one-shot, then restores the previous model.
    • Streams each reference model's full reasoning as labelled blocks before the aggregator's live-streamed synthesis, making ensemble deliberation visible.
    • New /learn <anything> command distills a reusable skill from a directory, URL, or recent workflow and writes it to the project's CONTRIBUTING.md automatically.
    • New /journey command in CLI and TUI displays a browsable, editable timeline of all memories and skills Hermes has accumulated, with delete support.

    +9 more show less

    • Adds a memory graph in the desktop app — a playable radial timeline of memories and skills that can be inspected and pruned.
    • New /goal completion contracts let users define evidence-based done criteria; the standing-goal loop judges completion by running actual project checks via a pre_verify hook.
    • Enables delegate_task to fan out multiple subagents in the background, consolidating all results into a single turn without blocking the chat.
    • New /prompt command opens $EDITOR for composing multi-line prompts in markdown, queuing the result as the next message.
    • Adds Google Vertex AI as a first-class provider, with automatic OAuth2OAuth2An 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. token minting and refresh from a service-account JSON or Application Default Credentials — no static key required.
    • Gateway supports scale-to-zero: goes dormant when idle, wakes on demand, and performs drain coordination before restarts or migrations to avoid dropping in-flight conversations.
    • Routes post-turn self-improvement (memory/skill save decisions) to a cheaper auxiliary model with digested context, reducing cost of the background learning loop.
    • Adds first-class coding Projects to the desktop app — a sidebar, coding rail, review pane, git worktree management, and a project → repo → lane model.
    • Security hardening: blocks cron base_url overrides that could exfiltrate provider credentials, locks down 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.-config persistence attack surfaceattack surfaceThe exposed systems, interfaces and access points an attacker could try to use. It includes internet-facing services as well as identities, software and other reachable resources., adds Slack xapp- token redaction, enforces cloud-metadata floor on every browser backend, and strips dynamic Hermes secrets from subprocess spawn environments.

    14 details in Everything

58. v2026.6.19 Jun 19, 2026 · issue 014

    Hermes v0.17.0 adds iMessage/Raft/WhatsApp channels, background subagents, image editing, Automation Blueprints, and a major memory tool upgrade.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.6.19 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.6.19

    └──▷ TRY IT
    Connect Hermes to iMessage through Photon's managed line pool — no Mac relay needed.
    copy

    $ hermes photon login

    Kick off a long research or build task in the background and keep the main session free while it runs.
    python copy

    delegate_task(background=true)

    Opt in 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.-powered skill consolidation while keeping routine curation token-free on all other runs.
    copy

    $ hermes curator run --consolidate

    • Adds iMessage support via Photon Spectrum — no Mac relay or BlueBubbles bridge required; authenticate with hermes photon login and a device code.
    • Adds Raft agent network integration as a new gateway channel, letting external Raft agents wake Hermes via a wake-channel bridge with metadata-only payloads.
    • Adds official WhatsApp Business Cloud API adapter — Meta's first-party hosted path with no QR-scanning bridge process.
    • Enables background/async subagents via delegate_task(background=true), returning a handle immediately so work continues in parallel.
    • Extends image_generate with image-to-image editing — pass a source image and prompt to edit, transform, or recompose it across all supported image providers.

    +13 more show less

    • Introduces Automation Blueprints — schedule automations by name through a guided form, slash command, or conversation instead of writing cron syntax.
    • Adds grok-composer-2.5-fast to the xAI OAuth model picker with full 200k context, accessible via an existing xAI Grok subscription.
    • Upgrades the memory tool with an operations array for atomic batch add/replace/remove edits evaluated against the final character budget in a single call.
    • Adds a full profile builder to the dashboard — configure model, skills, 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. servers from the browser without editing config.yaml.
    • Adds a security scan on each skill in the overhauled Skills Hub browser, alongside connected hubs, a Featured section, and full skill previews before install.
    • Introduces Curator cost optimization: LLM consolidation now requires explicit opt-in (curator.consolidate: true or hermes curator run --consolidate); routine pruning runs token-free by default.
    • Adds live subagent watch-windows in the desktop app, streaming a delegated agent's activity into its own pane.
    • Supports installing any VS Code Marketplace theme directly into the desktop app, with per-profile theme assignment.
    • Adds Telegram Bot API 10.1 rich message rendering — better formatting and native markup on by default, with opt-out.
    • Adds a pluggable CronScheduler provider system (cron.provider config key) with an InProcessCronScheduler and a Chronos NAS-mediated managed-cron provider for scale-to-zero scheduling.
    • Adds a signed-HTTP inbound relay receiver with connector-to-gateway channel auth and an enroll CLI.
    • Adds an official Unreal Engine 5.8 MCP server to the MCP catalog.
    • Enables finding and diffing user-modified bundled skills to surface local customizations.

    18 details in Everything

59. v2026.6.5 Jun 6, 2026 · release notes

    Hermes v2026.6.5 ships a native desktop app, browser admin panel, /undo, fuzzy model picker, and NVIDIA skills tap.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.6.5 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.6.5

    └──▷ TRY IT
    Back up a conversation that went off the rails and re-send your last message with edits, without losing the session.
    copy

    $ /undo 3

    Drop into the classic CLI interface for a single invocation even if your default is set to TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency..
    copy

    $ hermes chat --cli

    Run the Nous Portal quick-setup flow to go from a fresh install to a configured, chatting agent in seconds.
    copy

    $ hermes portal

    • Adds Hermes Desktop — a native Electron app for macOS, Linux, and Windows with streaming chat, drag-and-drop files, clipboard image paste, Cmd+K command palette, in-app self-update, and an inline model picker.
    • Enables connecting the desktop app to a remote Hermes gateway via OAuth or username/password, with per-profile remote host targeting and concurrent multi-profile sessions.
    • Adds a full browser-based administration panel: Channels page (Telegram, Discord, Slack, etc.), 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. catalog with enable/disable toggles, credential management, webhook creation, memory configuration, and one-click Debug Share.
    • Ships a complete Simplified Chinese (简体中文) translation across all desktop UI surfaces, switchable via Appearance settings (display.language).
    • Adds /undo [N] command to retract the last N user turns across CLI, TUI, and messaging platforms (Telegram, Discord, etc.), with the last message prefilled for editing.

    +7 more show less

    • Adds --cli flag to hermes chat and a configurable default interface (cli or tui), plus a unified /model command and Sessions overlay in the TUI.
    • Adds hermes portal as a human-readable alias for the Quick Setup flow via Nous Portal, getting new users from install to first message without docs.
    • Adds NVIDIA/skills as a built-in trusted Skills Hub tap alongside OpenAI, Anthropic, and HuggingFace — CUDACUDANVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first.-X, AIQ, cuOpt and other NVIDIA skills are one install away.
    • Adds fuzzy search to the model picker across desktop, web dashboard, TUI, and CLI; model catalog now refreshes hourly; multi-endpoint providers grouped to reduce duplicates.
    • Adds new models to the picker: deepseek-v4-flash, MiniMax-M3 (1M context), qwen3.7-plus.
    • Adds environments: relevance gate for skills, keeping context-specific skills (kanban, docker/s6) out of the index unless explicitly requested.
    • Enables pruning of unused built-in skills (not just agent-created ones) via the curator, with per-skill usage tracking.

    12 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The bundled skill set removes spotify, linear, kanban-codex-lane, debugging-hermes-tui-commands, and several empty category markers by default; setups relying on these skills being auto-loaded must install them explicitly via hermes skills install.
    • !Heavier/niche skills (dspy, subagent-driven-development, minecraft-modpack-server, pokemon-player, hermes-s6-container-supervision, and the Baoyu creative set) are no longer loaded by default and must be installed explicitly.
    • !Building the desktop app now requires Node ≥20.19 or ≥22.12.
60. v2026.5.29 May 29, 2026 · release notes

    Hermes v2026.5.29 adds explicit Docker insecure opt-in, full 19,932-entry skills catalog, and memory providers now receive completed-turn context.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.5.29 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.5.29

    └──▷ TRY IT
    Run Hermes in Docker with LAN access while explicitly opting into insecure dashboard mode — required after the bind-host inference removal.
    copy

    $ docker run -e HERMES_DASHBOARD_INSECURE=1 -p 0.0.0.0:8080:8080 nous-research/hermes-agent

    • Adds HERMES_DASHBOARD_INSECURE=1 env var as an explicit opt-in to disable the dashboard's loopback auth guard in Docker, replacing silent bind-host inference.
    • Expands the skills.sh catalog from 858 to 19,932 entries by walking the full sitemap instead of a single paginated page.
    • Memory providers now receive completed-turn message context, enabling richer recall across sessions.
    • Narrows hindsight recall_types default to observation-only, giving more targeted hindsight behavior out of the box.

    4 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Docker setups that previously relied on bind-host inference to enable --insecure must now explicitly set HERMES_DASHBOARD_INSECURE=1 or the dashboard's same-origin guard will remain active.
61. v2026.5.28 May 28, 2026 · release notes

    Hermes v2026.5.28 adds multi-agent Kanban swarms, promptware defense, Bitwarden secret management, 4500× faster session search, and a curated 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. catalog.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.5.28 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.5.28

    └──▷ TRY IT
    Spin up a full parallel agent swarm to decompose and execute a complex task across cheap and expensive models.
    copy

    $ hermes kanban swarm "Refactor the auth module and write tests"

    Discover and install a Nous-vetted MCP server interactively — credentials are prompted and written to ~/.hermes/.env automatically.
    copy

    $ hermes mcp

    • Adds hermes kanban swarm command to auto-decompose a task into a full Swarm v1 graph (root, parallel workers, gated verifier, gated synthesizer, shared blackboard) with per-task model overrides, scheduled start times, and retry fingerprinting.
    • Adds promptware/Brainworm-class injection defense: ~15 new threat patterns in tools/threat_patterns.py, memory scanning at load time, and delimiter markers on tool results to prevent impersonation of system content.
    • Rebuilds session_search into a no-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., three-mode tool (discovery ~20ms, scroll ~1ms, browse) — eliminating ~$0.30/call and ~90s latency.
    • Adds Bitwarden Secrets Manager integration: set BWS_ACCESS_TOKEN and Hermes pulls all provider credentials from Bitwarden at startup, with secrets.bitwarden.override_existing toggle and EU/self-hosted server support.
    • Adds ntfy as the 23rd messaging platform plugin — no signup or API key required, just a topic URL, enabling push notifications from cron jobs, kanban completions, or send_message.

    +9 more show less

    • Adds skill bundles: a single /<name> slash command loads a named group of skills simultaneously; ships three new optional skills: code-wiki, openhands, and web-pentest.
    • Adds TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. session orchestrator: list, switch, refresh, and close multiple live sessions within one Ink TUI window, with a session-scoped model picker.
    • Adds two new image generation providers: Krea 2 Medium ($0.03) and Krea 2 Large ($0.06), auto-discovered via hermes tools → Image Generation → Krea.
    • Adds hermes mcp interactive picker backed by a Nous-vetted MCP server catalog; installs servers with one keystroke and writes credentials to ~/.hermes/.env.
    • Adds OpenHands orchestration skill under optional-skills/autonomous-ai-agents/openhands/ for delegating coding sub-tasks to any LiteLLM-supported provider, usable in kanban swarms and /delegate flows.
    • Adds xAI Web Search as a plugins/web/xai/ provider alongside Brave/Tavily/Exa/SearXNG/DDGS/Firecrawl, reusing existing Grok OAuth or XAI_API_KEY credentials.
    • Cold-start improvements cut hermes --version latency 63% (701ms → 258ms), Termux cold start from 2.9s to 0.8s, and per-conversation function calls by 47% (399k → 213k for a 31-turn chat).
    • Adds worker status endpoints /workers/active, /runs/{id}, and /inspect for observing kanban swarm worker state.
    • Adds security-guidance plugin that pattern-matches dangerous code writes.

    14 details in Everything

62. v2026.5.16 May 16, 2026 · release notes

    Hermes v2026.5.16 adds SuperGrok OAuth, an OpenAI-compatible local proxy, X search, Teams integration, LINE/SimpleX platforms, and dozens of new capabilities.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.5.16 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.5.16

    └──▷ TRY IT
    Route Aider (or any OpenAI-compatible tool) through your Claude Pro or SuperGrok subscription — no API key needed.
    copy

    $ hermes proxy &
    aider --openai-api-base http://localhost:<port> --openai-api-key dummy

    Install Hermes from PyPI and launch immediately — no repo clone required.
    copy

    $ pip install hermes-agent && hermes

    • Adds xAI Grok as a SuperGrok OAuth provider with grok-4.3 bumped to a 1M-token context windowcontext windowThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. — no API key required.
    • New hermes proxy command exposes an OpenAI-compatible local endpoint backed by Claude Pro, ChatGPT Pro, or SuperGrok OAuth, letting tools like Codex, Aider, Cline, and Continue consume any subscription without an API key.
    • Adds x_search as a first-class built-in tool for searching X (Twitter) via OAuth or API key — no skill install or custom integration needed.
    • Adds full Microsoft Teams integration: Microsoft GraphMicrosoft GraphMicrosoft's unified REST API for accessing data and services across Microsoft 365, Azure AD, and related cloud products, giving cyber tools a single authenticated endpoint to query users, emails, files, and sign-in logs. auth, webhook listener, pipeline plugin runtime, and outbound delivery.
    • Hermes is now installable via pip install hermes-agent from PyPI — no repo clone or shell installer required.

    +18 more show less

    • Adds cross-session 1-hour Claude prompt caching (system prompt, skills, memory) for Anthropic, 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 Nous Portal backends.
    • Browser CDPCDPCisco Discovery Protocol, a Cisco-proprietary Layer 2 protocol that advertises device identity, capabilities, and topology between directly connected Cisco gear, giving cyber tools a passive way to map network infrastructure. calls are 180× faster via a persistent Chrome DevTools connection shared across browser_console evaluations.
    • Cold-start launch time reduced by ~19 seconds through deferred adapter loading, disk-cached model catalogs, parallel doctor checks, and a chat -q banner-skip flag.
    • Adds LINE and SimpleX Chat as first-class messaging platforms, bringing the total to 22.
    • /handoff now transfers the full active session — messages, tool calls, and context — live to a new model, persona, or profile.
    • Native platform button UI for clarify multiple-choice prompts on Telegram and Discord.
    • Discord channel history is now backfilled by default when the agent joins a channel or thread for the first time.
    • vision_analyze now passes raw pixels directly to vision-capable models (GPT-5, Claude, Gemini, Grok-vision) instead of converting to a text description.
    • Adds a per-turn file-mutation verifier footer summarizing paths, line counts, and deltas after every turn that writes or edits files.
    • Adds unified video_generate tool with pluggable provider backends, supporting new video models via a one-file plugin.
    • computer_use cua-driver backend now works with non-Anthropic models.
    • Heavy backends (Slack, Matrix, Feishu, DingTalk, image-gen SDKs, voice/TTS providers) now lazy-install on first use, with a supply-chain advisory checker scanning every install.
    • Adds Zed ACP Registry integration via uvx.
    • Adds native Windows beta support.
    • Adds 9 new optional skills.
    • Adds OpenRouter Pareto Code router.
    • Adds huggingface/skills as a trusted default skill tap.
    • Clickable URLs now render in any terminal.

    23 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The [all] extras no longer include dependencies now covered by lazy-installation — packages previously pulled in by pip install hermes-agent[all] may no longer be present at install time and will instead be fetched on first use.
63. v2026.5.7 May 7, 2026 · release notes

    Hermes v2026.5.7 ships durable multi-agent Kanban, /goal persistence, video analysis, 8 P0 security closures, and Google Chat as platform #20.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.5.7 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.5.7

    └──▷ TRY IT
    Lock the agent onto a long-running investigation target so it stays on task across multiple conversation turns without drifting.
    copy

    $ /goal Enumerate all exposed S3 buckets in the org and produce a findings report

    Archive stale curator entries and prune orphaned data in one pass to keep the knowledge store clean.
    copy

    $ hermes curator archive && hermes curator prune

    • Adds durable multi-agent Kanban board with heartbeats, zombie detection, per-task retries, hallucination gate, and multi-project support across profiles.
    • New /goal command locks the agent onto a target across turns (Ralph loop as a first-class primitive).
    • New video_analyze tool enables native video understanding on Gemini and compatible 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. models.
    • Adds xAI Custom Voices as a TTS provider with voice cloning support.
    • Adds Google Chat as the 20th supported messaging platform; introduces generic platform-plugin hooks so third-party adapters drop in without touching core.

    +22 more show less

    • Gateway now auto-resumes interrupted sessions after restart, /update restarts, or source-file reloads.
    • Security hardening: redaction is ON by default, Discord role-allowlists are now guild-scoped (closes CVSSCVSSCommon Vulnerability Scoring System, a method for describing a vulnerability's characteristics and severity. Its score is not the probability that attackers will exploit it. 8.1 cross-guild DM bypass), WhatsApp rejects strangers by default, TOCTOU windows closed across auth.json and 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., browser enforces cloud-metadata 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. floor, cron scans assembled skill content 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 hermes debug share redacts at upload.
    • Checkpoints v2 rewrites state persistence with real pruning, disk guardrails, and no orphan shadow repos.
    • New no_agent cron mode runs a script-only watchdog, skipping the agent entirely.
    • Adds allowed_channels / allowed_chats / allowed_rooms platform allowlist config across Slack, Telegram, Mattermost, Matrix, and DingTalk.
    • Providers are now pluggable via ProviderProfile ABC and plugins/model-providers/ directory.
    • API server gains X-Hermes-Session-Key header for stable per-session long-term memory.
    • 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. gains 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. transport with OAuth forwarding, stale-pipe retries, image results surfaced as MEDIA tags, and keepalive on long-lived lifecycle waits.
    • Curator grows hermes curator archive, prune, and list-archived subcommands; manual hermes curator run is now synchronous.
    • ACP gains /steer and /queue endpoints to direct the in-flight agent or queue follow-ups from Zed, VS Code, or JetBrains.
    • TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. adds /model picker with inline auth, collapsible startup banner sections, and a context-compression counter in the status bar.
    • Dashboard adds a Plugins management page, Profiles management page, sortable analytics tables, reverse-proxy support via X-Forwarded-Prefix, and a new default-large 18px theme.
    • SearXNG ships as a native search-only backend; web tools now support different backends per capability (search vs extract vs browse).
    • 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. response caching adds explicit cache control for models that expose it.
    • New [[as_document]] skill directive forces gateway to deliver output as a document on supported platforms.
    • New transform_llm_output plugin lifecycle hook lets plugins reshape or filter 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. output before it enters the conversation.
    • Nous OAuth tokens now persist across profiles via a shared token store.
    • QQBot gains native approval keyboards with feature parity to Telegram/Discord, plus chunked upload and quoted attachments.
    • Adds 6 new optional skills: Shopify (Admin + Storefront GraphQL), here.now, shop-app personal shopping assistant, Anthropic financial-services bundle, kanban-video-orchestrator, and searxng-search.
    • Adds new models: deepseek/deepseek-v4-pro, x-ai/grok-4.3, openrouter/owl-alpha (free), and tencent/hy3-preview.
    • Static gateway and CLI messages now translated to 7 locales: Chinese, Japanese, German, Spanish, French, Ukrainian, and Turkish.
    • Adds 100 new CLI startup tips covering cron, kanban, curator, plugins, and lesser-known flags.

    27 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Redaction is now ON by default — setups that relied on unredacted output in hermes debug share or logs will now receive redacted content without explicit opt-out.
    • !WhatsApp now rejects messages from unknown senders by default — existing deployments open to strangers will need to explicitly configure acceptance.
    • !Discord role-allowlists are now guild-scoped — cross-guild role grants that previously allowed DM access will no longer work.
64. v2026.4.30 Apr 30, 2026 · release notes

    Hermes v2026.4.30 adds autonomous skill curation, 5 new inference providers, 2 new messaging platforms, Spotify/Google Meet integrations, one-shot mode, and ~57% faster TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. cold start.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.30 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.30

    └──▷ TRY IT
    Run a quick, non-interactive prompt against a specific model without entering the TUI — useful in CI pipelines or shell scripts.
    copy

    $ hermes -z "Summarize the top 3 CVEs from today's NVD feed" --provider openrouter --model nous-hermes-3

    Check whether an update is available and back up your HERMES_HOME before applying it.
    copy

    $ hermes update --check

    Inspect which skills the Curator considers most- and least-used before manually pruning or after a curation cycle.
    copy

    $ hermes curator status

    • Adds hermes curator autonomous background agent that grades, prunes, and consolidates your skill library on a 7-day cron cycle, writing per-run reports to logs/curator/run.json and REPORT.md.
    • Adds hermes curator status to rank skills by usage, showing most-used and least-used entries.
    • Upgrades the self-improvement review fork to class-first (rubric-based) grading with active-update bias, proper provider/model/credential inheritance, and clean memory-provider shutdown.
    • Promotes ComfyUI v5 (official CLI + REST + hardware-gated local install) from optional to built-in by default.
    • Promotes TouchDesigner-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. from optional to built-in by default, expanded with GLSL, post-FX, audio, geometry, and 9 new reference docs.

    +20 more show less

    • Adds new Humanizer skill that strips AI-isms from text output.
    • Adds /reload-skills slash command and direct-URL skill install support.
    • Upgrades LM Studio from a custom-endpoint alias to a full native inference provider with dedicated auth, hermes doctor checks, reasoning transport, and live /models listing.
    • Adds four new inference providers: GMI Cloud, Azure AI Foundry (with auto-detection), MiniMax (OAuth with 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. browser flow), and Tencent Tokenhub.
    • Makes the gateway a plugin host with drop-in messaging adapters; ships Microsoft Teams as the first plugin-delivered platform.
    • Adds Tencent 元宝 (Yuanbao) as the 18th native gateway messaging platform with text and media delivery.
    • Adds native Spotify integration: 7 tools (play, search, queue, playlists, devices) behind PKCE OAuth, interactive setup wizard, bundled skill, and cron usage.
    • Adds Google Meet plugin for joining calls, transcribing, speaking, and following up via realtime OpenAI transport and a bundled Node bot server.
    • Adds hermes -z <prompt> non-interactive one-shot mode with --model, --provider, and HERMES_INFERENCE_MODEL support.
    • Adds hermes update --check preflight check and opt-in pre-update HERMES_HOME backup.
    • Adds a Models dashboard tab with per-model analytics and in-browser switching of main and auxiliary models.
    • Adds remote model catalog manifest so new 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 Nous Portal models appear without a release.
    • Adds native 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. image routing based on each model's actual vision capability rather than provider defaults.
    • Adds native multi-image sending across Telegram, Discord, Slack, Mattermost, Email, and Signal; adds centralized audio routing with FLAC support and Telegram document fallback.
    • Adds LaTeX rendering, /reload .env hot-reload, pluggable busy-indicator styles, opt-in auto-resume of last session, session delete from /resume picker with d, modified mouse-wheel line scroll, and a /mouse toggle to suppress ConPTY phantom mouse injection in the TUI.
    • Bundles Langfuse observability plugin and hermes-achievements plugin that scans full session history.
    • Adds a pluggable tts.providers.<name> registry and ships Piper as a native local TTS provider.
    • Adds Vercel Sandbox as an execute_code/terminal backend.
    • Adds configurable prompt_caching.cache_ttl (5-minute default, 1-hour opt-in) for cost savings on bursty sessions.
    • Cuts visible TUI cold start by ~57% via lazy agent init, lazy imports, mtime-cached load_config(), memoized get_tool_definitions() with 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.-cached check_fn results, and precompiled dangerous-command patterns.

    25 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Secret redaction is now off by default; working setups that relied on automatic secret redaction must explicitly opt in via redaction.enabled: true.
65. v2026.4.23 Apr 23, 2026 · release notes

    Hermes v2026.4.23 adds a full Ink TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency., native AWS Bedrock, 5 new inference providers, QQBot, and a vastly expanded plugin surface.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.23 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.23

    └──▷ TRY IT
    Launch the new interactive TUI for a rich, real-time agent session with streaming output and subagent observability.
    copy

    $ hermes --tui

    Course-correct a running agent mid-task without interrupting its turn or breaking prompt cache.
    copy

    $ /steer focus only on files under src/auth/ and skip any database migrations

    • Adds hermes --tui: a full React/Ink rewrite of the interactive CLI with sticky composer, live streaming, OSC-52 clipboard support, per-turn stopwatch, git branch status bar, /clear confirm, light-theme preset, and subagent spawn observability overlay.
    • Adds pluggable agent/transports/ layer with AnthropicTransport, ChatCompletionsTransport, ResponsesApiTransport, and BedrockTransport — each owning its own format conversion and API shape.
    • Adds native AWS Bedrock provider support via the Converse API.
    • Adds five new inference providers: NVIDIA NIM, Arcee AI, Step Plan, Google Gemini CLI OAuth, and Vercel ai-gateway (with pricing and dynamic discovery).
    • Adds GPT-5.5 access via ChatGPT Codex OAuth, with live model discovery wired into the model picker so new OpenAI releases appear without catalog updates.

    +18 more show less

    • Adds QQBot as the 17th supported messaging platform via QQ Official API v2, with QR scan-to-configure setup wizard, streaming cursor, emoji reactions, and DM/group policy gating.
    • Expands plugin surface: plugins can now register slash commands (register_command), dispatch tools directly (dispatch_tool), veto tool execution from pre_tool_call hooks, rewrite tool results (transform_tool_result), transform terminal output (transform_terminal_output), ship image_gen backends, and add custom dashboard tabs.
    • Adds /steer <prompt> command to inject mid-run nudges to a running agent after its next tool call without interrupting the turn or breaking prompt cache.
    • Adds shell hook support so any shell script can be wired as a Hermes lifecycle hook (pre_tool_call, post_tool_call, on_session_start, etc.) without writing a Python plugin.
    • Adds webhook direct-delivery mode to forward payloads straight to a platform chat without invoking the agent — enabling zero-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. push notifications for alerting and uptime checks.
    • Adds explicit orchestrator role for subagents with configurable max_spawn_depth, plus a file-coordination layer so concurrent sibling subagents share filesystem state without clobbering each other.
    • Adds a dedicated Configure auxiliary models screen in hermes model for per-task model overrides (compression, vision, session_search, title_generation).
    • Adds extensible web dashboard plugin system: third-party plugins can add custom tabs, widgets, and views; paired with live-switching themes controlling colors, fonts, layout, and density without a reload.
    • Adds dashboard i18n (English + Chinese), mobile-responsive layout, real per-session API call tracking, and one-click update and gateway restart buttons.
    • Adds per-provider and per-model request_timeout_seconds configuration.
    • Adds configurable API retry count via agent.api_max_retries.
    • Adds Kimi K2.6 across 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., Nous Portal, native Kimi, and HuggingFace.
    • Adds Xiaomi MiMo v2.5-pro and v2.5 on OpenRouter, Nous Portal, and native providers.
    • Adds Claude Opus 4.7 to the Nous Portal catalog.
    • Adds /model merging of models.dev entries for additional providers.
    • Adds Gemini routing through the native AI Studio API for improved performance.
    • Adds xAI Grok upgraded to the Responses API.
    • Adds Ollama cloud provider support, GLM continuation, think=false control, surrogate sanitization, and /v1 hint.

    23 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !auto auxiliary model routing now defaults to the main model for side tasks for all users — previously, aggregator users were silently routed to a cheap provider-side default.
66. v2026.4.16 Apr 16, 2026 · release notes

    Hermes v2026.4.16 adds Nous Tool Gateway: web search, image gen, TTS, and browser automation for Portal subscribers.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.16 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.16

    └──▷ TRY IT
    Enable gateway-backed tools for a Portal subscriber without configuring any separate API keys.
    copy

    $ hermes model

    Verify which gateway tools are active after opt-in via use_gateway config.
    copy

    $ hermes tools && hermes status

    • New Nous Tool Gateway gives paid Nous Portal subscribers automatic access to web search (Firecrawl), image generation (FAL / FLUX 2 Pro), text-to-speech (OpenAI TTS), and browser automation (Browser Use) — no extra API keys required.
    • Per-tool opt-in via use_gateway config key, with full visibility through hermes tools and hermes status.
    • Runtime now prefers the gateway over direct API keys when both are present.

    3 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The HERMES_ENABLE_NOUS_MANAGED_TOOLS environment variable is replaced by subscription-based detection via the Nous Tool Gateway; setups relying on that env var will no longer activate managed tools.
67. v2026.4.13 Apr 13, 2026 · release notes

    Hermes v2026.4.13 adds iMessage/WeChat, Fast Mode, local web dashboard, background process monitoring, backup/restore, and 16-platform coverage.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.13 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.13

    └──▷ TRY IT
    Enable priority-queue routing for low-latency responses when working with GPT-5.4 or Claude in an active session.
    copy

    $ /fast

    Watch a background build or server process and get notified the moment it prints a key event, without polling.
    yaml copy

    watch_patterns: ["listening on port", "ERROR", "build failed"]

    Snapshot your full Hermes setup before a major upgrade, then restore it on a new machine.
    copy

    $ hermes backup && hermes import <path>

    • Adds local browser-based web dashboard for managing agent settings, sessions, skills, and gateway configuration without editing config files.
    • Adds /fast toggle for priority-queue routing on OpenAI (GPT-5.4, Codex) and Anthropic (Claude) models for significantly lower latency.
    • Adds iMessage integration via BlueBubbles with auto-webhook registration and setup wizard.
    • Adds native WeChat (Weixin) support via iLink Bot API with streaming, media uploads, and markdown link handling.
    • Adds WeCom callback-mode adapter for self-built enterprise apps with atomic state persistence.

    +17 more show less

    • Adds Termux/Android support — run Hermes natively on Android with TUITUIA text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. optimizations, voice backend, and /image command.
    • Adds watch_patterns for background process monitoring, delivering real-time alerts when output matches defined patterns without polling.
    • Adds native xAI (Grok) provider with direct API access, model catalog, and setup wizard integration.
    • Adds Xiaomi MiMo as a first-class provider with setup wizard, model catalog, and empty response recovery.
    • Adds Qwen OAuth provider with portal request support.
    • Adds pluggable context engine slot via hermes plugins for swapping in custom context engines (filtering, summarization, domain injection).
    • Adds unified SOCKSSOCKSA proxy protocol that relays TCP (and optionally UDP) connections through an intermediary server, letting cyber tools route traffic through a pivot host or anonymizing layer without needing raw network access. proxy, DISCORD_PROXY, and system proxy auto-detection across all gateway platforms.
    • Adds hermes backup and hermes import commands for full backup and restore of configuration, sessions, skills, and memory.
    • Adds /debug slash command across all platforms and hermes debug share to upload a full debug report to a pastebin.
    • Adds /compress <focus> for guided context compression scoped to a specific topic.
    • Adds allowed-channels whitelist config for Discord.
    • Adds DISCORD_REPLY_TO_MODE setting for Discord.
    • Adds rate limit header capture displayed in /usage.
    • Adds custom providers to /model listings and resolution.
    • Adds fallback provider activation on repeated empty responses with user-visible status.
    • Adds comprehensive security hardening: path traversal protection in checkpoint manager, shell injection neutralization in sandbox writes, 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. redirect guards in Slack image uploads, Twilio webhook signature validation, API server auth enforcement, git argument injection prevention, and approval button authorization.
    • Expands platform support to 16 messaging platforms including BlueBubbles (iMessage) and WeChat alongside existing Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Email, SMSSMSShort Message Service, the protocol built into cellular networks for sending text messages between phones; cyber tools use it for alerting, two-factor authentication delivery, or phishing simulation over mobile channels., DingTalk, Feishu, WeCom, Mattermost, Home Assistant, and Webhooks.

    22 details in Everything

68. v2026.4.8 Apr 8, 2026 · release notes

    Hermes v2026.4.8 adds 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. 2.1, live model switching, background task notifications, Google AI Studio, and OSVOSVOpen Source Vulnerabilities, a vulnerability database and schema maintained by Google that gives cyber tools a machine-readable, cross-ecosystem format for querying and sharing CVE-style advisories. malware scanning.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.8 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.8

    └──▷ TRY IT
    Switch to a different model mid-session without restarting — useful when you need a more capable or cheaper model for the next task.
    copy

    $ /model gemini-2.5-pro

    Tail agent logs in real time to debug tool execution or provider errors across a running session.
    copy

    $ hermes logs --tail

    • Adds notify_on_complete for background tasks — agent is automatically notified when long-running processes (builds, tests, deployments) finish, eliminating polling.
    • New /model command enables live model and provider switching mid-session across CLI, Telegram, Discord, Slack, and other gateway platforms.
    • Adds native Google AI Studio (Gemini) provider with automatic context lengthcontext lengthThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. detection via models.dev registry.
    • 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. 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. 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. support for standards-compliant MCP server authentication.
    • Adds automatic OSV malware scanning of MCP extension packages on install.

    +20 more show less

    • Adds approval buttons on Slack and Telegram for dangerous command approvals — replaces typing /approve with native inline buttons.
    • Adds inactivity-based agent timeouts — agents are only timed out when truly idle, not when actively running tools.
    • Adds hermes logs command for tailing and filtering structured agent logs written to ~/.hermes/logs/ (agent.log + errors.log).
    • Adds config structure validation at startup to catch malformed YAML before it causes cryptic failures.
    • Expands plugin system: plugins can now register CLI subcommands, receive request-scoped API hooks with correlation IDs, prompt for required env vars during install, and hook into session lifecycle events.
    • Adds MiniMax TTS provider (speech-2.8) for text-to-speech.
    • Adds interactive model picker with inline buttons for Telegram and Discord.
    • Adds xAI (Grok) prompt caching via x-grok-conv-id header.
    • Adds model pricing display for 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 Nous Portal providers.
    • Adds free-tier Xiaomi MiMo v2 Pro model on Nous Portal for auxiliary tasks (compression, vision, summarization).
    • Adds Matrix Tier 1 support: reactions, read receipts, rich formatting, and room management.
    • Adds Discord channel controls and ignored-channels configuration.
    • Adds Signal MEDIA: tag delivery and Mattermost file attachments.
    • Self-optimized GPT/Codex tool-use guidance via automated behavioral benchmarking, patching 5 failure modes for improved reliability on OpenAI models.
    • Adds thinking-only prefillprefillThe first phase of inference, where the model reads the whole prompt in parallel before producing its first token. It drives time-to-first-token; decode is the token-by-token phase after it. continuation for structured reasoning on supported models.
    • Oversized tool results are now saved to file instead of being destructively truncated.
    • Adds HERMES_PORTAL_BASE_URL env var to override the Nous portal base URL during login.
    • Adds Z.AI endpoint auto-detection via probe and cache.
    • Adds non-agentic model warning when loading Hermes 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. models not designed for 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..
    • Security hardening: consolidated 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. protections, timing attack mitigations, tar traversal prevention, credential leakage guards, cron path traversal hardening, and cross-session isolation.

    25 details in Everything

69. v2026.4.3 Apr 3, 2026 · release notes

    Hermes v2026.4.3 adds pluggable memory backends, credential pool rotation, Camoufox stealth browser, secret exfiltration blocking, and ACP/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. editor integration.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.4.3 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.4.3

    └──▷ TRY IT
    Install the Camoufox stealth browser backend so the agent can browse without bot detection.
    copy

    $ hermes tools

    Configure multiple API keys for the same provider so Hermes rotates them automatically under load or on 401 failures.
    yaml copy

    # In config.yaml
    credential_pool:
      - provider: openai
        api_key: sk-key-one
      - provider: openai
        api_key: sk-key-two

    Maintain a persistent session across multiple API server requests so the agent retains context between calls.
    copy

    $ curl -X POST https://localhost:8080/v1/chat \
      -H 'X-Hermes-Session-Id: my-session-42' \
      -H 'Content-Type: application/json' \
      -d '{"message": "continue the analysis"}'

    • Adds pluggable memory provider interface — third-party backends (Honcho, 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., custom DBs) implement a provider ABC and register via the plugin system.
    • Adds same-provider credential pools with automatic least_used rotation and 401-triggered failover across multiple API keys via credential_pool config.
    • Adds Camoufox anti-detection browser backend for stealth browsing with persistent sessions, VNC URL discovery, and configurable 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. bypass; auto-install via hermes tools.
    • Adds inline diff previews for file write and patch operations in the tool activity feed.
    • Adds API server session continuity via X-Hermes-Session-Id headers and real-time tool progress streaming for Open WebUI integration.

    +11 more show less

    • Adds ACP support for client-provided MCP servers — VS Code, Zed, and JetBrains editor MCP servers are picked up as additional agent tools.
    • Adds secret exfiltration blocking: browser URLs and 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 are scanned for secret patterns, blocking URL-encoded, base64, 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. exfiltration attempts.
    • Expands credential directory protection to .docker, .azure, and .config/gh; redacts execute_code sandbox output.
    • Adds developer role support for GPT-5 and Codex models.
    • Adds Anthropic long-context tier 429 handling — automatically reduces context to 200k on tier limit hits.
    • Adds auto-detection of models from server probe during custom endpoint setup.
    • Adds skill-aware slash commands — gateway dynamically registers installed skills as slash commands with paginated /commands list.
    • Makes config.yaml the single source of truth for endpoint URLs, eliminating conflicts with environment variables.
    • Adds Honcho full integration parity as the reference memory provider plugin with profile-scoped host/peer resolution.
    • Adds token usage persistence for non-CLI sessions.
    • Adds DM thread sessions seeded with parent transcript to preserve context.

    16 details in Everything

70. v2026.3.30 Mar 30, 2026 · release notes

    Hermes v2026.3.30 adds multi-instance profiles, 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 mode, Feishu/WeCom/Slack multi-workspace support, fallback provider chains, and Exa search.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.3.30 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.3.30

    └──▷ TRY IT
    Run isolated staging and production agent instances from a single Hermes installation without credential collisions.
    copy

    $ hermes profile create staging
    hermes -p staging model set nous/hermes-3-70b
    hermes -p staging chat

    Keep agent tasks running even when your primary provider is rate-limited or down by configuring an automatic failover chain.
    yaml copy

    # config.yaml
    fallback_providers:
      - openrouter
      - nous-portal
      - local-ollama

    • New hermes profile create/list/switch/delete/export/import/rename commands enable fully isolated multi-instance deployments, each with its own config, memory, sessions, skills, and gateway service with token-lock credential isolation.
    • New hermes mcp serve exposes conversations, sessions, messages, and attachments to any MCP-compatible client (Claude Desktop, Cursor, VS Code) via stdio or Streamable HTTPStreamable HTTPAn MCP transport protocol that lets a client send JSON-RPC requests over HTTP POST while the server streams responses back using Server-Sent Events, replacing the older stdio and SSE transports for remote tool communication. transports.
    • Official Dockerfile enables running Hermes Agent in a container in both CLI and gateway modes with volume-mounted config.
    • New fallback_providers config key enables an ordered failover chain across multiple inference providers.
    • Adds full gateway adapter for Feishu/Lark with event subscriptions, message cards, group chat, image/file attachments, and interactive card callbacks.

    +12 more show less

    • Adds new gateway adapter for WeCom (Enterprise WeChat) supporting text, image, and voice messages, group chats, and callback verification.
    • Slack gateway now supports multi-workspace OAuth, connecting a single gateway to multiple workspaces via an OAuth token file with per-workspace bot token resolution.
    • Telegram adapter gains webhook mode as a production alternative to polling, plus configurable group mention gating (always, @mention-only, or regex-triggered).
    • Adds Exa as an alternative web search and content extraction backend alongside Firecrawl and DuckDuckGo, configured via EXA_API_KEY.
    • Skills and credential files can now be mounted into Modal and Docker remote backends so remote terminal sessions share the same skills and secrets as local execution.
    • Discord gateway adds a DISCORD_IGNORE_NO_MENTION env variable to skip messages that @mention other users or bots but not Hermes.
    • Discord gateway adds a processing reaction emoji that appears while a request is in flight and is removed on completion.
    • Matrix adapter sends voice messages as native Matrix voice events via MSC3245 instead of file attachments.
    • Mattermost adapter gains configurable mention behavior to respond without requiring an @mention.
    • Gemini 3.1 preview models added to 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 Nous Portal catalogs.
    • Rate limit (429) responses now display a user-friendly message with a Retry-After countdown.
    • Curated model lists are now preferred over live API probes when the probe returns fewer models.

    17 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !When no provider is configured, Hermes now raises a clear error instead of silently routing to OpenRouter.
71. v2026.3.28 Mar 28, 2026 · release notes

    Hermes v2026.3.28 adds Hugging Face inference, Telegram topic isolation, plugin lifecycle hooks, native Modal SDK, and a Nix flakeNix flakeA Nix packaging construct that pins all dependencies in a lockfile and exposes a standardized interface, giving cyber tools fully reproducible builds across any machine running the Nix package manager..

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.3.28 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.3.28

    • Adds Hugging Face Inference API as a first-class provider with auth, setup wizard, curated agentic model picker, and live /models endpoint probe.
    • Enables Telegram Private Chat Topics for project-based conversations with per-topic skill binding and isolated workflows within a single chat.
    • Activates plugin lifecycle hooks — pre_llm_call, post_llm_call, on_session_start, and on_session_end — now firing in the agent loop and CLI/gateway.
    • Replaces swe-rex dependency with native Modal SDK (Sandbox.create.aio + exec.aio), eliminating tunnel requirements for the Modal terminal backend.
    • Adds GPT_TOOL_USE_GUIDANCE to enforce tool calls in GPT models instead of prose descriptions, and strips stale budget warnings from conversation history.

    +15 more show less

    • Introduces Nix flake with full uv2nix build, NixOS module with persistent container mode, and auto-generated config keys from Python source.
    • Expands Nous Portal to 400+ available models via a single provider endpoint.
    • Adds /resume CLI handler and reopen_session API for returning to previous sessions.
    • Adds --source flag for third-party session isolation by origin.
    • Introduces session search recent-sessions mode — omit query to browse recent sessions with titles, previews, and timestamps.
    • Exposes compression.target_ratio, protect_last_n, and threshold in DEFAULT_CONFIG for ratio-based context compression tuning.
    • Persists reasoning across gateway session turns with new schema v6 columns (reasoning, reasoning_details, codex_reasoning_items).
    • Adds supply chain hardening: CI workflow scanning PRs for supply chain attack patterns, pinned dependency ranges, and regenerated uv.lock with hashes.
    • Adds per-model native output limits for Anthropic models (128K for Opus 4.6, 64K for Sonnet 4.6) replacing hardcoded 16K max_tokens.
    • Adds thinking-budget exhaustion detection to skip useless continuation retries when a model uses all output tokens on reasoning.
    • Gives subagents independent iteration budgets.
    • Surfaces all retry, fallback, and compression events to the user as formatted messages.
    • Adds auto-discovery of fallback IPs via DNS-over-HTTPSDNS-over-HTTPSA protocol that tunnels DNS queries inside encrypted HTTPS requests, hiding lookups from network observers. Cyber tools use it to evade DNS-based monitoring or filtering. when api.telegram.org is unreachable.
    • Removes /model slash command from CLI and gateway in favor of hermes model subcommand.
    • Increases API timeout default from 900s to 1800s to support slow-thinking models.

    20 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The /model slash command is removed from CLI and gateway; use the hermes model subcommand instead.
    • !The summary_target_tokens config key is replaced with ratio-based scaling; existing configurations referencing it will have no effect.
    • !The litellm dependency is removed; any external code or plugins that relied on it will break.
72. v2026.3.23 Mar 24, 2026 · release notes

    Hermes v2026.3.23 adds an OpenAI-compatible API server, 6 messaging adapters, 4 inference providers, and 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. 2.1 management.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.3.23 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.3.23

    └──▷ TRY IT
    Inject a local config file and a live URL into context before asking a question — useful for grounding the agent in current data without copy-pasting.
    copy

    $ hermes chat '@file config/rules.yaml @url https://example.com/policy "Does our config comply with this policy?"'

    • Adds OpenAI-compatible /v1/chat/completions API server with /api/jobs REST APIREST APIA web API design style where clients interact with resources over HTTP using standard methods like GET, POST, and DELETE, giving cyber tools a predictable, language-agnostic way to integrate with external services. for cron job management, SQLite-backed response persistence, input limits, field whitelists, and 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. origin protection.
    • Adds 6 new messaging platform adapters: Signal, DingTalk, SMSSMSShort Message Service, the protocol built into cellular networks for sending text messages between phones; cyber tools use it for alerting, two-factor authentication delivery, or phishing simulation over mobile channels. (Twilio), Mattermost, Matrix, and Webhook — joining existing Telegram, Discord, and WhatsApp support, with exponential-backoff auto-reconnect.
    • Adds @file and @url context injection with tab completions in the CLI for Claude Code-style context references.
    • Adds 4 new inference providers: GitHub Copilot (OAuth + token validation, 400k context), Alibaba Cloud / DashScope, Kilo Code, and OpenCode Zen/Go.
    • Adds hermes mcp command suite for installing, configuring, and authenticating 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 with full 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. 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. flow.

    +20 more show less

    • Adds gateway prompt caching — caches AIAgent instances per session to preserve Anthropic prompt cache across turns, reducing cost on long conversations.
    • Adds NeuTTS local TTS provider backend with built-in setup flow.
    • Enables CLI streaming by default with spinner and tool progress display during streaming mode.
    • Adds /statusbar command to toggle a persistent config bar showing model and provider info in the prompt.
    • Adds /queue command to queue prompts for the agent without interrupting the current run.
    • Adds /permission command to switch approval mode dynamically during a session.
    • Adds /browser command for interactive browser sessions from the CLI.
    • Adds /cost command for live pricing and usage tracking in gateway mode.
    • Adds /approve and /deny commands replacing bare text approval in gateway mode.
    • Adds show_reasoning config option to display reasoning/thinking blocks during inference.
    • Adds context pressure warnings for CLI and gateway modes.
    • Adds ${ENV_VAR} substitution support in config.yaml.
    • Adds real-time config reload — config.yaml changes apply without restarting the agent.
    • Adds custom_models.yaml for user-managed model additions.
    • Adds priority-based context file selection with CLAUDE.md support.
    • Adds merging of nested YAML sections instead of full replacement on config update.
    • Adds overhaul of context compression with structured summaries, iterative updates, token-budget tail protection, configurable summary endpoint, and fallback model support.
    • Adds context lengthcontext lengthThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. detection overhaul with models.dev integration, provider-aware resolution, fuzzy matching for custom endpoints, and /v1/props support for llama.cpp.
    • Adds eager fallback to a backup model on rate-limit errors.
    • Adds endpoint metadata support for querying local servers for actual context windowcontext windowThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. size and pricing.

    25 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !The ANTHROPIC_BASE_URL environment variable is removed to avoid collisions — configurations relying on it will no longer apply.
73. v2026.3.17 Mar 17, 2026 · release notes

    Hermes v2026.3.17 adds streaming, a plugin architecture, native Anthropic provider, voice mode, 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, persistent shell, and CDPCDPCisco Discovery Protocol, a Cisco-proprietary Layer 2 protocol that advertises device identity, capabilities, and topology between directly connected Cisco gear, giving cyber tools a passive way to map network infrastructure. browser attach.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.3.17 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.3.17

    └──▷ TRY IT
    Extend Hermes with a custom security tool (e.g., a CVE lookup) without forking the project.
    copy

    $ cp my_cve_lookup_tool.py ~/.hermes/plugins/

    Attach Hermes browser tools to an already-open Chrome session for live inspection and interaction via CDP.
    copy

    $ /browser connect

    • Adds unified streaming infrastructure delivering real-time token-by-token output in the CLI and all gateway platforms.
    • Introduces a first-class plugin architecture: drop Python files into ~/.hermes/plugins/ to add custom tools, commands, and hooks without forking.
    • Adds native Anthropic provider with Claude Code credential auto-discovery, OAuth 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, and native prompt caching — no 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. required.
    • New smart approvals system learns which commands are safe and remembers preferences; /stop command immediately kills the current agent run.
    • Adds Honcho memory integration with async writes, configurable recall modes, session title support, and multi-user isolation in gateway mode.

    +16 more show less

    • Adds voice mode: push-to-talk in CLI, voice notes in Telegram/Discord, Discord voice channel support, and local Whisper transcription via faster-whisper.
    • Concurrent tool execution via ThreadPoolExecutor runs multiple independent tool calls in parallel, reducing multi-tool turn latency.
    • New privacy.redact_pii setting automatically scrubs personally identifiable information before sending context 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. providers.
    • New /browser connect command attaches browser tools to a live Chrome instance via Chrome DevTools Protocol.
    • Adds Vercel AI Gateway as a provider, routing Hermes through Vercel's model catalog and infrastructure.
    • Rebuilt centralized provider router with call_llm API, unified /model command, auto-detect provider on model switch, and direct endpoint overrides.
    • New ACP server enables VS Code, Zed, and JetBrains to connect to Hermes as an agent backend with full slash command support.
    • Persistent shell mode lets local and SSH terminal backends maintain shell state (working directory, env vars, aliases) across tool calls.
    • Adds --pass-session-id flag to include the session ID in the system prompt.
    • Adds Anthropic Context Editing API support.
    • New system gateway service mode allows running Hermes as a system-level systemd service.
    • Reasoning hot reload allows changing reasoning settings in gateway mode without a restart.
    • Adds Agentic On-Policy Distillation (OPD) RL training environment for distilling agent policies.
    • Native Anthropic auxiliary vision support uses Claude's native vision API for image processing instead of OpenAI-compatible endpoints.
    • Accepts unlisted models with a warning instead of rejecting them outright.
    • CLI token counts are now persisted to the session DB for use with /insights.

    21 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Default group gateway sessions now use per-user isolation instead of shared state across users in group chats.
74. v2026.3.12 Mar 12, 2026 · release notes

    Hermes Agent v2026.3.12 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. client support, multi-platform messaging gateways, git worktree isolation, filesystem rollback, and 70+ bundled skills.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v2026.3.12 https://github.com/NousResearch/hermes-agent.git
    # already have the repo? check out this version:
    $ git checkout v2026.3.12

    └──▷ TRY IT
    Run an isolated agent session in a git worktree to safely work on a feature branch without touching your main working tree.
    copy

    $ hermes -w

    Send a one-shot query non-interactively for use in scripts or CI pipelines.
    copy

    $ hermes --quiet "Summarize the open CVEs in SECURITY.md"

    Check the health of all configured 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. providers and integrations before deploying.
    copy

    $ hermes doctor

    • Adds MCP (Model Context ProtocolModel Context ProtocolModel 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.) client with stdio and HTTP transports, reconnection, resource/prompt discovery, and server-initiated LLM sampling requests.
    • Adds multi-platform messaging gateway supporting Telegram, Discord, Slack, WhatsApp, Signal, Email (IMAP/SMTPSMTPThe application-layer protocol used to send email between servers and from clients to servers; cyber tools use it to relay alerts, phishing simulations, or exfiltrated data over standard mail infrastructure.), and Home Assistant with unified session management and media attachments.
    • Adds hermes -w flag to launch isolated agent sessions in git worktrees for safe parallel work on the same repository.
    • Adds filesystem checkpoints with automatic snapshots before destructive operations and /rollback command to restore prior state.
    • Adds Skills Ecosystem with 70+ bundled and optional skills across 15+ categories, a Skills Hub for community discovery, per-platform enable/disable, and prerequisite validation.

    +48 more show less

    • Adds ACP Server for VS Code, Zed, and JetBrains editor integration via the Agent Communication Protocol standard.
    • Adds data-driven CLI skin/theme engine with 7 built-in skins (default, ares, mono, slate, poseidon, sisyphus, charizard) and support for custom YAML skins.
    • Adds centralized provider router with unified call_llm() / async_call_llm() API and resolve_provider_client() for automatic credential resolution across all LLM consumers.
    • Adds support for OpenAI Codex (Responses API) with ChatGPT subscription and OAuth vision/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. content.
    • Adds Kimi Code API, z.ai/GLM, Kimi/Moonshot, MiniMax, and Azure OpenAI as first-class providers.
    • Adds self-hosted Firecrawl support.
    • Adds 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 routing via provider_preferences configuration.
    • Adds /reasoning command for configuring reasoning effort level and display toggle.
    • Adds /personality command for setting and disabling custom agent personalities.
    • Adds /verbose slash command to toggle debug output at runtime.
    • Adds /insights command for usage analytics, cost estimation, and activity patterns.
    • Adds /background command for managing background processes.
    • Adds /compress, /usage, and /update slash commands to the gateway.
    • Adds /resume command for switching to named sessions in the gateway.
    • Adds session naming with unique titles, auto-lineage, rich listing, and resume-by-name.
    • Adds interactive session browser with search filtering.
    • Adds Honcho AI-native cross-session user modeling.
    • Adds --quiet / -Q flag for programmatic single-query mode.
    • Adds --fuck-it-ship-it flag to bypass all approval prompts.
    • Adds hermes tools subcommand with a curses UI for per-platform tool enable/disable.
    • Adds hermes doctor subcommand for health checks across all configured providers.
    • Adds hermes update subcommand with auto-restart for the gateway service.
    • Adds user-defined quick commands that bypass the agent loop.
    • Adds simple fallback model configuration for provider resilience.
    • Adds configurable subagent provider/model with full credential resolution.
    • Adds tool call repair middleware with auto-lowercase and invalid-tool handling.
    • Adds smart context lengthcontext lengthThe maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. probing with persistent caching and banner display.
    • Adds Nous Portal as a first-class provider in the setup wizard.
    • Adds config migration system (currently v7) and modular setup wizard with section subcommands.
    • Adds terminal bell notification when the agent finishes a task.
    • Adds clipboard image paste via Alt+V / Ctrl+V.
    • Adds up/down arrow history navigation in the interactive CLI.
    • Adds Signal messenger gateway via signal-cli-rest-api with media URL support.
    • Adds full email gateway platform via IMAP/SMTP.
    • Adds Home Assistant REST tools and 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. gateway with service discovery.
    • Adds DISCORD_ALLOW_BOTS config option for bot message filtering.
    • Adds WhatsApp native media sending (images, videos, documents) and multi-user session isolation.
    • Adds Telegram forum topic session isolation and browser screenshot sharing via MEDIA: protocol.
    • Adds configurable background process watcher notifications.
    • Adds edit_message() for Telegram, Discord, and Slack with fallback.
    • Adds /model validation against the live API instead of hardcoded lists.
    • Adds /model and /provider unified single view.
    • Adds reasoning effort configuration via /reasoning command.
    • Adds iteration budget pressure via tool result injection and shared budget across parent and subagent delegation.
    • Adds atomic write for .env to prevent API key loss on crash.
    • Adds multiple named custom provider support.
    • Adds loading indicators for slow slash commands.
    • Adds tools summary flag.

    53 details in Everything