I built Agent Ready Kit, a small open-source CLI for one question that keeps showing up in AI-assisted development:
Can an AI coding agent actually understand and change this repo without wasting the first 20 minutes guessing?
The tool audits a repository locally and gives it an agent-readiness score. It checks the boring-but-important things that make agents useful instead of chaotic:
- clear README and install/run instructions
- test command discovery
- CI/workflow setup
- license and contribution basics
- env examples
- issue templates
- docs/examples
- AGENTS.md or other agent instructions
It can also generate a starter AGENTS.md so the next coding session has a better map of the project.
Why I made it
A lot of repos are technically open source but not very agent-friendly. Humans can infer missing setup steps from experience; agents usually burn tokens, make guesses, or ask for context that could have been written down once.
I wanted a quick pre-flight check before handing a repo to Codex, Claude Code, Cursor, Copilot, Gemini CLI, or any other coding agent.
Try it
Repo: https://github.com/subhanA-UA/agent-ready-kit
Live page: https://subhana-ua.github.io/agent-ready-kit/
Example:
npx agent-ready-kit
What I want feedback on
If you use AI coding agents in real repos, I would love feedback on what the score should include next.
The next checks I am considering:
- package manager lockfile consistency
- flaky or missing test detection
- stale docs vs current scripts
- repo-specific agent safety rules
- PR-review readiness
If this saves you a setup pass, or if the scoring misses something important, tell me. That feedback is more useful than a vague launch post.
Top comments (1)
Checking for a test command is useful, but the stronger signal is whether it succeeds from a clean checkout with only the documented setup steps. I'd make the local audit distinguish "present" from "verified" across the README, environment example, CI workflow, and generated AGENTS.md, because stale instructions can give an agent more confidence while sending it in the wrong direction. Lockfile consistency fits that model too: a smaller score grounded in executable evidence will be more trustworthy than a broad checklist that rewards files merely existing.