AI Policy Read - read an open-source repository's AI contribution policy before you (or your agent) contribute.
aipr fetches the governance files that usually carry AI rules
(CONTRIBUTING.md, AI_POLICY.md, AGENTS.md, CLAUDE.md, ...), classifies
the repository's stance with weighted phrase matching, and answers one
question: can an AI-assisted or autonomous contribution land here?
$ aipr asciimoo/hister
aipr: asciimoo/hister
[BLOCKED] human-only policy
sources: CONTRIBUTING.md, README.md
confidence: 1.0 score: 19.0
autonomous contribution: NOT SAFE - require human co-authorship
[+5.0] ...Issues and PR descriptions must be fully human-written...
[+5.0] ...AI should never be the main author of the PR...
$ aipr apache/maka
aipr: apache/maka
[UNKNOWN] no explicit AI policy found
exit=2
More repositories are publishing explicit AI policies - from "we welcome
AI-assisted work" to "agents are strictly forbidden". Violating one burns the
contributor (and, for autonomous agents, the operator): rejected PRs at best,
blocks at worst. aipr makes the check mechanical and cheap, for humans
deciding where to spend review effort and for agents deciding where to spend
their quota.
pip install git+https://github.com/yunaremaia/aipr.git
# requires Python 3.10+; GH_TOKEN recommended (anonymous API calls rate-limit fast)
export GH_TOKEN=ghp_xxx # classic token with public repo read accessNo dependencies beyond the standard library. pytest only to develop.
aipr OWNER/REPO # classify a GitHub repository
aipr --text FILE # classify a local governance file
aipr --json OWNER/REPO # machine-readable output| Verdict | Meaning | Autonomous-safe? |
|---|---|---|
human_only |
AI must not be the main author / human-written only / bans agents | no |
restrictive |
heavy process: mandatory disclosure + human-in-the-loop requirements | no |
disclose_ok |
allowed with a disclosure trailer (Assisted-by: AI) |
yes* |
permissive |
explicitly welcomes AI-assisted contributions | yes |
unknown |
no explicit policy found | ask first |
* still follow the disclosure rules - "safe" means no human co-authorship required by policy, not no obligations.
| Code | Meaning |
|---|---|
| 0 | all inspected repos are autonomous-safe |
| 1 | at least one repo is restricted or human-only |
| 2 | at least one repo is unknown / no policy found (ranks worse than 1) |
| 64 | usage error |
Batch mode: aipr owner/repo1 owner/repo2 ... prints one block per repo
(JSON array with --json) and the exit code reflects the worst result —
so an unverified repo can never pass a gate silently.
Policy fetches are cached on disk for 24h (~/.cache/aipr, configurable via
AIPR_CACHE_DIR / AIPR_CACHE_TTL), so repeated scans cost zero API calls.
Use --no-cache to force a fresh fetch.
Weighted regex matching over concatenated governance text. Restrictive phrases score positive ("must be fully human-written" +5), permissive ones negative ("we warmly welcome AI-assisted" -3.5). The strongest signals force the verdict; weak mixed signals lean restrictive on purpose - when in doubt, do not send a bot.
Known limits: English-only patterns; phrase matching cannot understand nuance; a repo can carry policy in unusual files we don't probe. Treat UNKNOWN as "read it yourself".
Early beta - battle-tested against a handful of real policies (hister, modular, polars, MDAnalysis, maka). Rule additions welcome: open an issue with the policy text and the verdict you expected.
MIT