AI marks

How this was made

Methodology

Where the data comes from, what each mark means, and what it cannot tell you.

Who did the research

Everything on this site was researched and written by an AI agent: Claude (Anthropic), working in Claude Code sessions. The agent found the projects, read the READMEs, docs, issues and source files linked from each claim, wrote the project files, and built the collector and this site. No person has checked the claims against their sources, read the code behind a feature mark, or tried the products. The maintainer reviews the site as a whole, not entry by entry. Read every mark as the agent's reading of the linked source, follow the link before relying on it, and please open an issue when it is wrong.

Sources and dates

Each project has a file in data/projects/, written by the agent: features, platforms and AI policy, each claim with a link (a file at a commit, a docs page, an issue) and the date the agent read it. Projects were found through the matrix.org ecosystem lists and the rooms of the #community:matrix.org space, plus one GitHub topic per category (matrix-client, matrix-homeserver, matrix-bot, matrix-bridge), keyword searches on GitHub and jryans/awesome-matrix. Other GitHub topics, topics on other forges, other awesome lists and This Week in Matrix have not been searched yet, so projects are missing. All entries were researched on 5 Nov 2026. Project logos are trademarks of their projects and are shown only to identify each project; each comes from the project's own repository or site (the source is in the project's file under icon) and is served from this site. Projects without one get a monogram.

Development metrics are computed by a collector from each repository's git history (a blobless clone, so every forge gets the same numbers) plus the forge's API for stars, issues, the archived flag and languages (GitHub, GitLab, Forgejo and Gitea; SourceHut and others give blanks). Lines of code come from scc on a shallow checkout. The current metrics were collected on 5 Nov 2026. Nothing from a scanned repository is ever run.

“Not found” is never the same as “no”: it means the agent looked and found nothing either way. It is drawn with a dashed outline.

The three AI axes

They are shown separately and never merged into one score.

AI policy

What the project says about AI-made contributions.

From CONTRIBUTING, codes of conduct, PR templates, the README, docs, pinned issues and maintainers' own statements, quoted with a link. Restricted means allowed with disclosure or limits.

AI in development

Whether the developers use AI tools on the repository.

Evidence: agent instruction files (CLAUDE.md, AGENTS.md, .cursorrules, .github/copilot-instructions.md and others), commits whose trailers or “Generated with” statements credit an AI tool, AI agent bot authors, and CI workflows that run AI actions or review bots, counted over the last 365 days. A trailer counts only when it names an AI: a tool's own address (noreply@anthropic.com, Copilot's GitHub address, …) or the tool's name with model words (“Claude Opus 4.5”, “GitHub Copilot”), so a co-author who shares a name with a tool (Claude, Devin, Jules) is not counted. An LLM API key in CI is not counted either: it usually tests the product's own AI feature. Config only: agent files but no attributed commits. Occasional: some attributed commits but under 5 % of last year's (the rubric says 1 to 4 %; a project with even one attributed commit below 1 % is also counted here, with its share shown), or AI review or agent workflows. Regular: 5 % or more, and at least three attributed commits (one AI commit out of twelve is not a habit). Attribution is a lower bound: most AI use leaves no trace in a repository, so “none found” does not mean none was used.

The collector suggests a level; the mark shown is the one the agent set in the project's file, compared with that suggestion. Where they differ the project page says why: an AGENTS.md that only tells agents to stay away is not AI use, and a project that rewrote its history to strip AI trailers, or says in prose that it is written with AI, is marked from its own statement. Merged pull requests that disclose AI use (squash merges drop the trailers) were searched on GitHub for the bot, tool and bridge lanes and are linked as evidence.

AI in the product

Whether the product itself has AI features.

Features of the product that call a language model or other generative AI, from docs and code. Optional means the user turns it on or can turn it off. On-device machine learning that generates nothing (a background blur in calls) is not counted, and is mentioned in the entry's note.

Funders' AI policies

Some projects are paid for by grants whose funder has its own rules on generative AI (NLnet's GenAI policy, for example, which in version 2.0 bans AI-produced artefacts in funded work). Those are shown on the project page as a separate panel, never merged into the project's own policy, because they bind the funded work rather than every contributor.

NLnet's policy has three versions: 1.0 (from 2025-12-08) and 1.1 (from 2026-01-26) allow generative AI with disclosure but do not pay for purely generated work; 2.0 (from 2026-10-20) allows only read-only use (review, finding flaws), with disclosure. Each grant gets the version that its dates and NLnet's FAQ point to: grantees in NGI Zero programmes already running may keep following 1.1, and a grant that ended before 2025-12-08 predates every version and is marked “none found”, with the reason. The panel gives the grant's dates, so they can be read next to the commit history; the site draws no conclusion from the two together.

Other funders checked on 2026-11-05: Outreachy “strongly discourages” generative AI and leaves the rule to each mentor; Google Summer of Code leaves it to each organisation. Both bind interns and contributors rather than a project's grant, so they are not shown per project. No rule on generative AI was found for the Sovereign Tech Agency, the Prototype Fund or the Open Technology Fund.

The AI-wary and AI-friendly switch

The switch in the header only changes how the marks are drawn. AI-wary draws AI use (and policies that allow it) with a red cross and “none found” with a green check; AI-friendly does the reverse. Restricted policies get an amber half mark in both; an unknown policy stays grey. Every mark keeps its icon and text, so colour is never the only signal. Nothing is hidden or reordered unless you filter. Neutral is the default; your choice stays on this device.

Activity status

Commits are counted on the default branch, without merge commits, by the date they landed. Authors are de-duplicated by email and name, honouring .mailmap. For projects with several repositories (an Android and an iOS app, say), the status uses all of them together. A mirror or an old home that shares most of its history with another repository is counted once, and a library the project builds on (listed with the role upstream) is shown but left out of the project's numbers. A change that landed on the branch twice (rebased or cherry-picked again, with the same author, author date and subject) counts once. For forks that merge their upstream (SchildiChat, Rooster, Vommet, Sable's first version) only the fork's own line of history is counted (git log --first-parent), so the upstream's work is not credited to the fork. The bus factor is the fewest authors who made half of last year's commits. Issue response is the median time to the first comment by someone other than the reporter, for issues opened in the last 180 days (where the forge allows reading it).

The bot filter

Commits by these bots count neither as commits nor as authors: dependabot, renovate, renovate-bot, weblate, hosted weblate, github-actions, gitlab-bot, forgejo-actions, pre-commit-ci, mergify, snyk-bot, greenkeeper, imgbot, allcontributors, semantic-release-bot, release-please, codecov, deepsource-autofix, transifex, crowdin, translate bot, translations bot, localazy, pyup-bot, depfu, stale, copybara, any author with [bot] in the name, and translation-platform commits (“Translated using Weblate”).

AI agents that commit under their own name (aider, amp, claude, coderabbit, codex, copilot, cursor, devin, gemini, jules, openhands, sweep) are different: their commits count as changes and as AI evidence, but not as authors.

Caveats

Corrections

Spotted something wrong or missing? Open an issue with a link to the source. Data is under CC BY 4.0, code under MIT. The whole dataset is one file: data.json.

This project's own AI use

In keeping with what it asks of others: this site was researched and built with Claude (Anthropic) in Claude Code, with people directing the work and merging it. Every commit names the model that wrote it in a Co-Authored-By line.

In this site's own terms: policy encouraged, development use regular, no AI in the product (the pages call no AI service).