Analytics

Each distinct posting observed in the displayed period is mapped to a product area from its title and department. About 17% of observed roles name an area specifically enough to place. Other titles remain unclassified. A documented product-area mention describes that posting; it does not measure the vendor's total investment.

Incomplete. Only about 17% of open roles carry a title specific enough to name a product area; the rest are generic engineering and go-to-market titles counted as "other". Read the placed roles as a floor on where a vendor is investing, never as the whole picture.

The identified postings in each vendor's latest successful board capture, with the actual source date beside each vendor. Observed empty boards count as zero; missing, stale or gapped boards are excluded from rankings. These are current board observations, not a total over a month or a headcount. Three completed comparable calendar months are required for a trend.

Incomplete. The displayed period contains successful captures in 3 distinct weeks. Short collection history limits trend comparisons. Failed captures retain labelled prior observations; postings may stay up after a role is filled.

Distinct postings observed in each month across covered vendors, split by the function inferred from the title. Postings describe advertised roles, not completed hires or a company's spending. Name companies above and this chart compares them one against another.

Incomplete. Job titles are not standardised across companies; the same title can describe different functions. In the latest displayed month, about 10% of postings could not be classified.

A posting can qualify for several regions, so these counts overlap and cannot be added as jobs. Shares use unique identified postings. Remote scope separates explicit worldwide, region restricted and unspecified eligibility; it does not identify offices.

Incomplete. A remote role says where a company will hire, not where it has an office, and about 11% of roles carry a location we cannot place.

Revenue growth plus free cash flow margin, each company measured over its own most recently completed fiscal year. The dashed line is the rule itself — growth and margin trading off along it is normal; a company sitting well above it is doing both at once.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Quarterly revenue as filed. Check Point and CyberArk are foreign private issuers that file one annual report rather than quarterly ones, so their series is drawn as yearly markers, not a line — a line through the empty quarters between them would read as data we don't have. A series reaching an "ended" marker has stopped filing, which means an acquisition or delisting, not a company whose revenue collapsed to zero.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Filed operating expense lines as shares of revenue within each exact fiscal period. Combined SG&A is a separate segment. Companies have different fiscal dates and are unranked; expense labels do not establish engineering effort or product value.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Full source-item comparisons require current source hashes, extraction and matching versions, and independent semantic review. Unverified historical classifications remain a separate ledger. Fiscal-year labels are not equal reporting periods, and unaccepted comparisons produce no sector narrative.

Incomplete. Risk-factor parsing and year-to-year heading matching need review. Complete extraction does not establish that an addition, removal or rewording is correct. Per-company coverage separates extracted filings from comparisons computed from the current inputs; the matching model still needs an evaluated sample before this chart can earn Solid.

Share of analysed repositories with an AI-attributed commit or an assistant configuration file observed in repository history. The two signals overlap; neither measures current usage or how much code was written with assistance. How this is measured →

For scale: GitHub went from 1 billion commits in all of 2025 to 275 million a week by April 2026, on pace for 14 billion this year if growth held, per GitHub's COO. That's his straight-line math, not a GitHub forecast.

Incomplete. 258 repos show commits with AI attribution and 108 have an assistant config file in their recorded history. This artifact has not yet passed the analysed-cohort validation. The signals overlap; neither measures current assistant usage or the share of code written with assistance.

Repositories carrying a committed config file that names a specific assistant, cumulative by the month that file first appears. Commit attribution is a separate measure and is never added to this one. The dotted line is AGENTS.md, the generic convention any agent can read, kept apart from the named assistants.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Tools in the watchlist by primary language, cumulative by the year the earliest retained root commit occurred, which may be imported history. The current analysed primary language is assigned retrospectively to this current watchlist. This is not historical language adoption, product founding, usage or popularity.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

The direct dependencies the most tools have in common, one ecosystem at a time. Counted once per tool, from each repo's own manifests — direct entries only, not the transitive closure.

Browse all 2,228 shared packages → and see which tools declare each one

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Published releases and tag events in the last 12 months, shown separately and alphabetically. Continuous observation duration and observed per-month rates are unknown. This retained-history view mostly covers public repositories, so — of the — tools here are open-source.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

How well each tool documents what it ships. Every release note is scored on depth, usage examples and transparency (breaking changes, advisories); a tool is scored on its median note out of its recent ones, so the three parts of each bar always add up to its total. Each tool is judged on its best captured channel — the release itself, a CHANGELOG in the repo, or a changelog page. An undiscovered channel can still change its score; a zero describes the captured text, not proof that the project publishes no documentation.

Incomplete. 1 tracked tool has a zero score for its captured text. A zero does not prove that release documentation is absent. An undiscovered changelog or an empty captured release body can change the ranking; channel discovery is still incomplete.

Captured service incidents from each vendor's public status page, listed alphabetically with their observation coverage. Reported incident time is the union of captured incident intervals over the last 90 days, including incidents limited to one component. It does not measure platform-wide downtime or uptime. Comparisons require a verified common monitoring window and component scope. Maintenance notices are excluded; bars show the worst reported severity. Click a bar to inspect its primary source.

Incomplete. Reported incident intervals may concern one component or an unspecified part of a service. Missing durations and uneven archive coverage limit comparison; global uptime cannot be calculated from these records.

Published releases per category over completed calendar months. Current observed-repository rates use a fixed matched numerator and denominator. Counts describe release activity among tracked tools, not investment, headcount, or whether a problem has been solved.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Compare observed spans only within identical endpoint sources and date precision. Year-only launch records retain bounds. Recent evidence does not verify operation, and closure remains unknown.

Incomplete. Commercial start dates come from a fallback chain ending in a hand-curated launch year, not repo history, so the commercial bars carry considerably more uncertainty than the open-source ones beside them.

Category averages of each tool's normalized signal index. The source declares ISO weeks or measured capture intervals; 100 is each signal's highest observed index in that history. Net counter changes can include removals. These indices describe signal shape, not audience size or adoption.

Incomplete. Download and star counters have only been collected since August 2026, so most tools' earlier weeks rest on Hacker News mentions alone — which under-count commercial products with no public repo or package.

Compare the same observed signals across 26 baseline and four recent periods. The source declares ISO weeks or measured capture intervals and their timing below. Net counter changes describe gains and removals, not gross actions, unique people, adoption or a cause of growth. Hacker News alone is shown separately when available.

Incomplete. Download and star counters have only been collected since August 2026, so most tools' earlier weeks rest on Hacker News mentions alone — which under-count commercial products with no public repo or package.

Tool classes documented in source-bound public reports, by half-year. Reports can describe several incidents and tools. Missing sources and classifications remain unknown; these observations do not establish incident prevalence.

Tools documented in collected reports, last 12 months

Incomplete. Captured reports cover only incidents publishers chose to describe. Missing sources and unevaluated tool labels remain unknown; these observations do not establish incident prevalence or a shift in attacker behavior.

Verified repository-creation and Show HN submission counts from a declared topic/keyword query cohort. The newest completed month at capture requires source evidence no older than 35 days. Historical counts keep their original query-capture dates; every partition still needs complete proof, and missing proof produces gaps. Partial observations stay separate, and source dates are not product launch dates.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

How much each tool's machine-facing surface changed in the last year — OpenAPI endpoints and MCP server tools added or removed. Additions are a count of observed expansion; removals may break integrations and need checking against the source. Surfaces without comparable captures are unavailable.

Incomplete. API: 40 tracked tools have comparable captures. MCP: no comparable captures in the displayed period. Coverage is limited to valid captured surfaces. Missing history is unavailable, not zero changes; changes between captures can be missed.

Share of classified commits labelled new capability (feat:) or fixes (fix:), from the last year of commits across tracked open-source tools. Commit counts do not measure effort.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Classifier matches for AI, agent features or MCP support in captured release notes, among tracked tools that released each quarter. Tools that are AI products themselves are excluded. A release that only fixes or mentions AI does not count. How this is measured →

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Share of note-bearing releases whose text matches each theme, per quarter. Matches include fixes and incidental mentions, not just shipped features.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Separate date-recency views of tracked repositories. Missing dates, contributor counts and capture times stay unknown. Contributor counts use each provider’s counting method, not distinct people across projects. First archived release/tag dates are not product launch dates.

Don't trust yet. This chart conflicts with its retained source evidence. Its comparisons are withheld while the conflict is resolved.

Each category's release volume last quarter against its own trailing four-quarter median — a ratio, so a big category can't out-rank a small one on size alone. Above 1× is accelerating; below is cooling.

† Comparison withheld for categories without sufficient comparable evidence.

We can't compare these categories with their own past yet, so this shows how many releases we recorded in each one last quarter. It doesn't say which ones are speeding up.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Every bug-fix bullet we can pull out of release notes, sorted into twelve failure families. This counts what vendors write down, so it describes disclosed defects — not all of them.

Incomplete. Only fix lines matching the keyword taxonomy enter a failure family. Unmatched text can change both the counts and the ranking; these are classified mentions, not a complete defect inventory.

The observed share of matched fix-note lines assigned to one family. Unresolved samples are unknown; alphabetical tool order avoids an unsupported ranking. Classifications still need an independent evaluation.

Incomplete. Built on the same partial classification, and only 425 of 633 tools clear the minimum sample. A tool missing from this chart is not a clean one — it is one we cannot yet say anything reliable about.

Distinct failure families against total defect volume. Variety rises with size on its own, so the tools worth a second look are the ones sitting well off the trend.

Incomplete. Variety is bounded by what the rules can recognise, so a tool failing in ways we do not classify looks narrower than it is. Read a tool against others of similar volume, never against the whole field.

What share of each quarter's release-note bullets describe a fix. This reads release-note prose; the work-breakdown chart reads commit prefixes. They read different records of the same tools and are not expected to agree.

Incomplete. Measures the share of release-note bullets that describe a fix, which rewards a vendor for writing less as much as for shipping fewer defects. It also disagrees with the work-breakdown chart above by design — different source, different coverage.

How often a failure family is fixed again within three releases or ninety days. Repetition is not proof the earlier fix failed — an actively developed subsystem gets fixed repeatedly for ordinary reasons.

Incomplete. The noisiest chart on this page. It counts a family being fixed again soon after, which is not evidence the earlier fix failed — an actively developed subsystem is fixed repeatedly for ordinary reasons. Read it as repetition, nothing stronger.

Projects that had a steady release rhythm and have since stopped — each measured against its own cadence, not a fixed threshold, so an annual release cycle isn't mistaken for silence. Most recently gone quiet first. This is not a judgement about a project or its maintainers: a finished tool legitimately stops shipping, and a repo can be actively developed without cutting releases. Read the cadence and the gap together, and check the repo before drawing a conclusion.

Incomplete. The source observations and calculation for this chart have not yet passed the current evidence checks.

Provisional links count normalized authored identities with qualifying, non-overlapping commits in both captured histories. Exact email and account declarations are not confirmed people. Unknown histories are separate from analysed isolates; shared authorship does not establish maintenance or affiliation. Inspect source commits and capture coverage below.

Incomplete. 797 of the repos we analyse appear on this map; the rest either share no contributors with anything else we track or have unparsed history. Absence from the map is not evidence that a tool stands alone.