Agent Friendly Code

Changelog

What each release shipped from the roadmap.

0.9.1Which file each agent reads2026-10-21
  • `/agents` — which repo files each AI coding agent reads, per its own docs, and how many leaderboard repos have each.
  • AGENTS.md vs CLAUDE.md — `/check/agents-md/vs-claude-md` says which to use, which agents read each, and what Claude Code does when a repo has both. The AGENTS.md check page lists what to put in the file and what to leave out.
  • The AGENTS.md check uses the first of AGENTS.md, CLAUDE.md and AGENT.md with content, so an empty AGENTS.md no longer hides a full CLAUDE.md. `docker-compose.yaml` and `.github/CONTRIBUTING.adoc` now count; a directory named like a checked file no longer does.
  • Live Score pages get the repo page's section menu, agent files panel, file links under All checks, and the pull-request Action snippet.
  • Repo page titles name the host, descriptions lead with the repo's own, and "Same topics, any language" links up to six repos that share its topics.
  • Every page's footer lists every agent board, language, check and topic.
  • Visitors from the EEA, UK and Switzerland are asked before Google Analytics sets cookies.
0.9.0Repo pages that answer for every agent2026-10-16
  • Agent boards — `/model/<id>` ranks every repo for one agent and opens with what that agent reads: its four weightiest checks, their pass rates, and how many repos have its own file.
  • Repo pages show the repo's description and topics, which agent files it has (linked to the file), each agent's score with the repo's rank on that agent's board, and how it compares with the leaderboard and its language.
  • AGENTS.md examples — `/check/agents-md/examples` lists every AGENTS.md and CLAUDE.md on the leaderboard, linked to the file, with its length, stars and score.
  • Topic pages — `/topic/<slug>` ranks repos sharing a topic once enough of them do; `/topics` lists them. Every repo is on a language page too, with small languages under `/language/other`.
  • Score your repo from the home page — the Live Score form sits in the home hero, and a result names the agent the repo is best set up for.
  • Check details read the same everywhere: "14,688 characters", "2 GitHub Actions workflows", and check descriptions in plain language. Scores are unchanged.
  • Pages fit a phone screen without sideways scrolling, and an unexpected error offers "Try again" instead of a bare error page.
0.8.0Browse repos by language2026-10-10
  • Language hubs — `/language` lists every language, and `/language/<slug>` ranks that language's repos and shows the checks they most often miss. Similar repos on a repo page are now the ones scoring closest to it, not the language's top three.
  • Readable repo URLs — a repo page now lives at `/repo/<host>/<owner>/<name>` instead of `/repo/:id`. Old numeric links redirect, and the address works in any letter case.
  • Renamed repos keep their page — when a tracked repo is renamed or moves to another owner, its old page address redirects to the new one, and old badge and `/api/score` URLs keep answering.
  • The AGENTS.md / CLAUDE.md check no longer passes on Cursor rule files. `.cursor/rules` and `.cursorrules` count only under the Cursor rules check, as `/methodology` describes.
  • Repos dropped from the tracked list leave the leaderboard at the next rescore, and a weekly check flags tracked repos that were renamed, archived, deleted or made private.
  • The pull-request workflow shown on a repo page targets that repo's default branch, and only appears for GitHub repos.
0.7.0Score any repo on the fly2026-08-25
  • On-demand scoring at `/score` — paste any public GitHub repository URL and get its full score in about a second, even if the leaderboard has never indexed it. Same signals, same per-model breakdown, same suggestions as an indexed repo page. Already-indexed repos redirect to their canonical repo page.
  • Scored without cloning — the repo is reconstructed from GitHub's tree API rather than `git clone`, so a score costs one API call and a handful of small file fetches. Results are cached for an hour, and nothing about a scored repo is stored.
  • Recent scores on `/score` — the repos you have scored are kept in your own browser and listed for one-click return. They are never sent anywhere.
  • Every live score carries the commit it was computed from, so a cached result says exactly what it reflects.
  • The home page now tells you once, on your first visit after a release, what that release shipped — then stays quiet until the next one.
  • Kimi CLI is the ninth scored agent — its own weight profile on `/methodology`, leaderboard view, and `?model=kimi-cli` badge.
0.6.0Auto-refresh — scores stay fresh on their own2026-05-20
  • Scheduled rescoring — a 6-hourly GitHub Actions cron re-runs `bun run seed` over every tracked repo, commits the refreshed `data/rank.db` to `main`, and Vercel redeploys. Scores stay current without anyone re-running the seed by hand. The job skips the commit when no `overall_score` actually changed.
  • Score-delta on the repo page — `/repo/:id` shows a `+N.N` / `-N.N` diff (green/red) under the overall score so a regression since the last rescore is visible at a glance. Hover or focus reveals a score-history popover with the previous value, the change, and the absolute scored-at date. Hidden on a repo's first scoring.
  • 16 new seed repos in the curated set so the cron has more material to refresh on each cycle.
0.5.0Scorer ships outside the dashboard2026-05-02
  • PR score-diff GitHub Action — `hsnice16/agent-friendly-action` on the GitHub Marketplace. Scores base + head locally inside CI and posts a single PR comment with the overall delta + per-signal breakdown. Vendored scorer (works even if the dashboard is down); opt-in via `AGENTS_BADGE_TOKEN`. Repo detail pages now show a copy-paste workflow snippet under "Catch score regressions on every PR".
  • Portable agent skill — `npx skills add hsnice16/agent-friendly-skill#v0` installs into any vercel-labs/skills-compatible host (Claude Code, Codex, Cursor, Cline, Copilot, Continue, Roo Code, …). Scores the user's current repo locally with a vendored scorer and recommends a model — produces identical output regardless of which host invokes it.
  • Dedicated integration pages on the dashboard — `/action` (workflow snippet + Marketplace link + FAQ) and `/skill` (install command, score → model mapping, optional `SessionStart` hook snippets for Claude Code and Codex).
  • Public `GET /api/score?host=<host>&repo=<owner>/<name>` — natural lookup endpoint for any tool that has a repo URL but not the dashboard's internal id. Returns `{ repo, signals, modelScores }` on 200, `{ error: "not_indexed" }` on 404. Built for external integrators; the action and the skill score locally.
0.4.0Credible scores + better discoverability2026-04-27
  • Per-model rationales now cite each agent's official documentation — `/methodology` renders a Sources panel under every model, and the 'illustrative weights' framing is gone. Reweights: Devin CI 1.0 → 0.7, Codex AGENTS.md 0.7 → 0.9, Cursor AGENTS.md 0.6 → 0.8.
  • Four new agent-specific instruction-file signals — `.cursor/rules/*.mdc` (Cursor), `GEMINI.md` (Gemini CLI), `.openhands/setup.sh` (OpenHands), `.aider.conf.yml` (Aider). Each is a per-agent boost on the model that reads it.
  • `/about` page — independent project, no vendor affiliation, links from the footer; gives the methodology a human face for both readers and LLM crawlers.
  • `/llms.txt` markdown manifest — Perplexity / Claude / ChatGPT-search can ingest the site's pages and public API endpoints cleanly.
  • Per-repo Open Graph images — sharing a `/repo/:id` URL renders a 1200×630 preview with the repo + score instead of the default site card.
0.3.0Embeddable scores + broader coverage2026-04-25
  • Embeddable score badges — `/api/badge/<host>/<owner>/<name>.svg` returns an SVG you can drop into a README; per-model variants via `?model=<id>`. Repo detail pages show a copy-paste markdown snippet.
  • Package lookup — `/package` page (registry dropdown + name input) and `/api/package/<registry>/<name>` JSON resolve npm / PyPI / Cargo packages to their source repo's score; unscored ones get a pre-filled GitHub-issue link to request scoring.
  • Same-language alternatives — repo detail pages show up to 3 alternative repos (same host + same language) ranked for the selected model.
  • Broader agent coverage — Gemini CLI, Aider, OpenHands, and Pi added to the per-model leaderboard with illustrative weights, flagged as such on `/methodology`.
  • Smoother leaderboard navigation — `ScoreBar` fills animate between pages instead of popping in.
0.2.0Complete the dogfood2026-04-24
  • This project scores ≥90 on its own rubric — overall 90.6, with each per-model score within 3 pts of the overall.
  • Leaderboard rows are fully clickable — click anywhere in a row to open the repo detail page (the external open link still goes to the host).
  • Curated set expanded ~5× — ~125 repos now ranked across JS/TS, Python, Rust, Go, JVM, C/C++, Swift, Ruby, Dart, language runtimes, AI-native models, and coding agents (up from ~25 at 0.1.0).
  • Desktop leaderboard shows 32 rows per page (mobile keeps 16), with pagination moved below the list so the toolbar stays clean.
  • Floating back-to-top button appears in the bottom-right when you're near the end of a long page.
  • Leaderboard footer shows the total number of tracked repos and when they were last scored.
0.1.0First public release2026-04-22
  • Public leaderboard ranking open-source repos by agent-friendliness across GitHub, GitLab, and Bitbucket.
  • Per-model scoring — separate ranking for Claude Code, Cursor, Devin, and GPT-5 Codex, each with its own signal weights.
  • Repo detail page — score breakdown per signal and the top improvements ranked by score-gain for the selected model.
  • Methodology page explaining what's measured today and what isn't.
  • Leaderboard controls — live search, host filter (GitHub / GitLab / Bitbucket), rank + stars sort, paginated results.
  • Light + dark theme that follows the OS preference; mobile-friendly layout down to 320-px viewports.
  • JSON API — `GET /api/repos` and `GET /api/repo/:id` for programmatic access.
  • MIT licensed — free to reuse, adapt, and self-host.