agents-md-lint

Your CLAUDE.md is lying to your agent. agents-md-lint finds the paths and commands your context file claims exist — and no longer do.

⭐ Star us on GitHub if your CLAUDE.md still describes a module you deleted.

pipx install agents-md-lint && agents-md-lint check

CLAUDE.md: 5 of 6 references broken
  CLAUDE.md:14  path: 'src/proxy/' — path does not exist
  CLAUDE.md:18  path: 'src/main.rs' — path does not exist
    did you mean: crates/tokenmiser/src/main.rs

That output is real. It is TokenMiser, whose context file described a src/ layout for months after the code moved to a Rust workspace under crates/.

Why this exists

Every agent workflow depends on a context file, and the failure mode is the dangerous kind: the agent doesn't error, it answers. Fluently, confidently, from a description of your repo that stopped being true three refactors ago.

Nobody notices, because nothing breaks loudly. The file just quietly stops matching the code.

What it checks

Non-zero exit past a threshold, so it drops into CI or a pre-commit hook:

- run: pipx install agents-md-lint && agents-md-lint check

Design stance

Cheap and deterministic. Path resolution, PATH lookup, git ls-files. No LLM call, no network, no embedding pass, zero runtime dependencies. A staleness check that costs an API call will not be run every session, and a check nobody runs is worse than none.

False positives are the failure mode. A linter that flags prose gets uninstalled after one run, so the extractor is deliberately conservative and most of the test suite is cases it must stay quiet about — URLs, globs, ~/paths, <placeholders>, $VARS, system paths, prose in backticks, ” ```python ” blocks, shell comments, and a project's own CLI (which is rarely on PATH during development).

Two of those were found by running agents-md-lint on its own CLAUDE.md.

Reports, never rebuilds. Fixing the file is your job; it knows what you meant, and this tool doesn't.

Measured on real repositories

Run across the Open Intelligence Labs monorepo:

Every one of the 30 findings was verified by hand against the filesystem. The two most interesting: a Rust project documenting a src/ layout that had become crates/, and a project listing a console/ directory that was never built.

Most of the rest are scaffolds whose context files describe a planned layout. That is arguably not "drift" — it is a file written ahead of the code — but the tool cannot tell intent from absence, and reporting it is the honest behaviour.

Install

pipx install agents-md-lint     # or: uv tool install agents-md-lint

agents-md-lint check               # every context file under .
agents-md-lint check CLAUDE.md     # one file
agents-md-lint check --json        # machine-readable
agents-md-lint check --threshold 3 # tolerate up to 3 findings

Honest limits

This checks two things. Paths and commands. It does not check whether your architecture description is still accurate, whether a documented convention is still followed, or whether the advice is any good — those need judgement, and judgement needs an LLM, which is explicitly out of scope.

Unquoted references are not checked. "edit the main.py file" is invisible to it. Backticks are the signal that something is a reference rather than prose, and without that signal the false-positive rate is unacceptable.

Commands are resolved, never run. Whether pytest exists is checkable; whether it passes is not, and running commands out of a context file would be a code-execution hazard.

The category ceiling is low. Developer tooling of this kind tops out in the low thousands of stars. This exists because it is useful and finishable, not because it is a growth play.

Roadmap

• Path and command checks with move suggestions
• JSON output, thresholds, CI exit codes
• Code index / knowledge graph staleness (commits since build)
• Generated-doc staleness (source changed after doc written)
• Embedding-cache staleness ratio
• Pre-commit hook

Part of the Open Intelligence Labs ecosystem

• TokenMiser — LLM cost control
• actants — agent SDK

License

MIT