Are We AI Friendly?

Where open source projects stand on AI: AI-assisted code, AI-found issues and AI-written text. Every rating comes from the project's own published policy, with quotes.

A fully static Astro site, served as static assets on Cloudflare Workers. There is no server code, database or API.

How it works

GitHub issue ("Add a project")
  └─ maintainer adds "ready-for-review" label (or the submitter or a maintainer comments)
       └─ intake workflow: fetch submitted sources → research agent finds policy pages
                           → AI review + classifier → quote check → comment on the issue
            ├─ passes → snapshot + data file → pull request → human review + merge
            └─ needs changes → "needs-changes" label → submitter edits or comments → review runs again
                                                                                        └─ deploy workflow → Cloudflare
weekly check workflow: re-fetch sources → unchanged text? skip : re-classify → pull request

• data/projects/<slug>.json is the source of truth: per-area verdicts, conditions, quotes and sources.
• snapshots/<slug>/*.txt holds the plain text of every source, so policy changes show up as git diffs. The hash in each source is the sha256 of its snapshot.
• Stances are computed, never stored or picked by the model (src/lib/stance.ts). One scale, used per area and overall: Friendly, Mostly friendly, Unfriendly, plus No policy. Per area: banned or minor help only → Unfriendly; disclosure or other extra rules → Mostly friendly; allowed or only baseline rules (accountability, human review, licence) → Friendly; not addressed → No policy. Overall follows code, capped at Mostly friendly when issues or text are Unfriendly.
• The model only fills in the schema. It reads each source document in full, every quote it returns must appear word for word in the snapshot (scripts/lib/quotes.ts), and anything doubtful is listed in the PR for the reviewer.
• Research agent (scripts/lib/research.ts). Before classification, a tool-calling loop looks for the project's AI rules. It reads the submitted pages, the issue's "Anything else?" field and comments by the submitter or maintainers (OWNER, MEMBER, COLLABORATOR), follows links, lists GitHub repository folders and searches the web through Firecrawl's keyless search API. It only picks pages, at most 5 on top of the submitted ones, each with a reason; it never rates. Pages may be on any domain, but must be official: published by the project, its core team or foundation. Checks on top of the agent's own judgment:
  • Social media, link aggregators, Q&A sites and encyclopedias are never accepted (NEVER_SOURCES).
  • The script tracks how each page connects to the project: on its own site or repository (homepage domain and subdomains, repository owner, hosts of submitted pages), linked from one of those pages (or on a site they link to), or linked by a person in the issue. A page with no such link gets one request to confirm it, and if picked anyway is flagged for the maintainer.
  • The classifier judges every page's publisher. A page the agent found by itself that it calls third-party or not a policy is dropped and the project is classified again without it, so nothing on the site comes from it.
  • When nothing mentions AI the agent picks the main contributing guide, so the project is listed as "No policy". Tool calls, steps and fetched text are capped. Submitting sources is optional; if the agent finds nothing either, the submitter is asked for links.
• Submission review (scripts/lib/review.ts). The model also says, per source, who published it and whether it is a policy. Problems the submitter can fix (form errors, unreadable links, already tracked, third-party or non-policy sources they gave, no policy pages found at all) get a "Changes needed" comment and the needs-changes label. Doubts about the AI output (low confidence, unclear publisher, a doubtful page the research agent found by itself, dropped quotes, disagreement with the submitter's own reading) still open the PR, labelled needs-review. The model never sees the submitter's reading, so it can't anchor on it.
• Re-runs. The workflow removes ready-for-review when it starts, so a maintainer can add it again at any time. While an issue has needs-changes, editing its body starts a new review on its own; a comment by the submitter or a maintainer does so at any time. Both stop after MAX_REVIEWS_ON_EDIT (5) review comments per issue; after that a maintainer re-adds the label. A re-run updates the existing pull request and closes outdated ones.

Styling

Tailwind CSS 4 through the official @tailwindcss/vite plugin. Typeface: Recursive, self-hosted via @fontsource-variable/recursive (its casual axis is used for headings via the casual utility).

• src/styles/global.css holds the design tokens in @theme: neutral white and grey (GitHub-style), near-black text, a blue accent for focus and hover, and one colour per mood. Tailwind's default palette is cleared, so only site colours exist (bg-paper, text-muted, border-line, ...). Dark mode swaps the same variables.
• Each stance is shown as a face on a square tile (src/components/MoodTile.astro, drawings in src/lib/faces.ts), both for the project and for each area. Labels: Friendly, Mostly friendly, Unfriendly, No policy; STANCE_INFO[...].plain keeps the literal wording ("AI-friendly with conditions") for titles, llms.txt and agents.
• Mood colours come from the tone-* utilities, which set --tone; components use bg-(--tone). The class map lives in src/lib/tones.ts.
• Utilities go in the markup with Astro's class:list. Interactive states use built-in variants: group-open: and has-open: for rows, aria-pressed: for the friendliness filters, aria-[current=page]: for navigation. Layout is mobile-first.

Development

Requires Bun.

bun install
bun run dev          # local site at http://localhost:4321
bun run build        # static site in dist/
bun run preview      # serve dist/ with wrangler, like production
bun test             # unit tests
bun run typecheck
bun run validate     # schema, snapshot hashes, quotes found in snapshots

Classifier settings

Any OpenAI-compatible chat completions endpoint works. Copy .env.example to .env (Bun loads it automatically):

Variable	Meaning
AI_BASE_URL	e.g. https://api.openai.com/v1, https://openrouter.ai/api/v1, http://localhost:11434/v1
AI_API_KEY	API key for that endpoint
AI_MODEL	model name
AI_RESPONSE_FORMAT	json_schema (default, strict structured output) or json_object for providers without schema support
AI_MAX_INPUT_CHARS	safety cap on text sent per project (default 300000); truncation is flagged for review

bun scripts/intake.ts --issue 3               # review a GitHub issue with its comments (repo: --repo or $GH_REPO)
bun scripts/intake.ts --body-file issue.md   # review an issue-form body; writes review-comment.md (and pr-body.md if it passes)
bun scripts/check.ts                          # check all sources (only calls the model if text changed)
bun scripts/check.ts --force curl             # re-classify one project
bun scripts/snapshot.ts <slug> <url>...       # snapshot sources for a hand-written entry

Hand-written entries use "classifiedBy": "manual"; classified ones use "ai". Run bun run validate before committing.

Setup checklist

Cloudflare

1. Create an account API token (Manage Account → Account API Tokens) with Account: Workers Scripts Edit and Account Settings Read for the account, and Zone: Workers Routes Edit, Zone Read and DNS Edit for the site's zone.
2. Add GitHub secrets CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID.
3. Add the domain as a zone in the same account. routes in wrangler.jsonc attaches it to the Worker as a custom domain on deploy.
4. Keep AI crawlers allowed: in the zone's Security → Bots settings, leave "Block AI bots" and the managed robots.txt off. Being readable by agents is the point of llms.txt.

GitHub

1. Settings → Actions → General: allow GitHub Actions to create pull requests.
2. Secrets: AI_BASE_URL, AI_API_KEY, AI_MODEL (optional AI_RESPONSE_FORMAT). Optional variable: AI_MAX_INPUT_CHARS. The provider settings are secrets so the public Actions logs never show which AI provider is used; GitHub masks secret values in logs. For the same reason classifier errors leave out the URL, host and provider response, and data files record "classifiedBy": "ai" rather than the model name.
3. Labels: submission, ready-for-review, correction (needs-changes and needs-review are created automatically). Issue forms skip labels that don't exist yet.
4. Pull requests opened by the workflows use GITHUB_TOKEN, so they don't trigger the Validate workflow. The intake workflow validates and builds before opening its PR, and the deploy workflow validates again before deploying, so invalid data never ships.

For AI agents

/llms.txt, /llms-full.txt, /projects.json, and /projects/<slug>.md for each project.