Blog
Which AI code reviewers read AGENTS.md and CLAUDE.md?
Published: November 8, 2026
Three review tools now read a repo's agent instructions without being told to: CodeRabbit reads both AGENTS.md and CLAUDE.md, GitHub Copilot code review reads AGENTS.md, and Anthropic's Code Review reads CLAUDE.md. Cursor Bugbot reads its own BUGBOT.md and, by its docs, not your Cursor rules. The difference that matters more than which file is read is which branch it is read from, because a pull request that edits the rules can change what it is checked against. Each fact below is from the vendor's docs, read on 8 November 2026.
Which code reviewer reads which rules file?
| Reviewer | AGENTS.md | CLAUDE.md | Its own file | Rules read from |
|---|---|---|---|---|
| CodeRabbit | Yes, by default | Yes, by default | Also Cursor, Copilot, Gemini, Windsurf and Cline files; .coderabbit.yaml path instructions | Not documented |
| GitHub Copilot code review | Yes | Not documented | .github/copilot-instructions.md and .instructions.md files | The head branch (the pull request) |
| Claude Code Review (Team and Enterprise) | Not documented | Yes | REVIEW.md | REVIEW.md from the base or default branch, never the pull request |
| Cursor Bugbot | Not documented | Not documented | .cursor/BUGBOT.md; .cursor/rules/*.mdc do not apply | Not documented |
| Claude in GitHub Actions | If your prompt says so | Yes | Your workflow and prompt | Whatever your workflow checks out |
| Harbor | Yes | Yes | Also Cursor and Copilot rules and CONTRIBUTING.md; on private repos, the team's rules | The base branch, before the pull request |
The Bugbot row is the one that surprises teams. Its docs say plainly that “Cursor project rules (*.mdc files in .cursor/rules/) do not apply to Bugbot runs,” so a team on Cursor keeps one rule set for the agent and a second, in BUGBOT.md, for the review. A .md file linked from BUGBOT.md does count, which is the cheapest way to point it at AGENTS.md.
Why does the branch matter?
GitHub's docs say Copilot code review “reads repository custom instructions, agent instructions, and agent skills from the head branch.” That is the convenient choice for a pull request that improves the instructions: the review uses the new version at once. It also means a pull request can soften a rule and break it in the same diff, and be reviewed against the softened rule.
That is not a far-fetched case. Coding agents edit AGENTS.md and CLAUDE.md as part of their work, often to record what they just did. Anthropic made the opposite choice for REVIEW.md, reading it from the base or default branch “and never from the pull request's own changes.” Harbor does the same for every rules file.
To test your own reviewer, open a pull request that deletes one rule from AGENTS.md and breaks that rule in the code. If the review stays quiet, it read the head branch.
What can a review check, and what belongs somewhere else?
A review is the slowest and least certain place to catch a broken rule. Use it for the rules nothing earlier can see.
| Kind of rule | Example | Best place to enforce it |
|---|---|---|
| Formatting, imports, naming | Named exports only | A linter or formatter, in CI |
| A command that must never run | No force push to main | A hook or a branch rule |
| A design decision | Retry provider webhooks on 5xx only, never 4xx | Pull request review |
| Where code goes | Billing logic stays in packages/billing | Pull request review |
| A rule that is out of date | An old auth library the repo already removed | Delete it; a reviewer will enforce it faithfully |
A review also needs rules it can check. “Write clean code” gives it nothing to cite; “Never retry on 4xx” does. Check your rule files flags the lines no one can check.
How do I check pull requests against AGENTS.md for free?
Run Claude in your own CI. Anthropic's GitHub Actions guide sets it up from Claude Code with /install-github-app, on your runners and your API key, and says to put review criteria in CLAUDE.md. If your rules live in AGENTS.md, add one line to CLAUDE.md so both are read:
# CLAUDE.md @AGENTS.md
Then ask the review, in its prompt, to quote the rule behind each comment. A finding that cannot name its rule is an opinion, and the author will treat it as one. Two limits remain. You pay for every run in API usage, and the review checks the file at whatever commit the workflow checked out, which for a pull request is usually its head.
What do rule-file reviewers still miss?
- Rules that never reached the file. Most team rules are settled in a review thread or a Slack channel. A reviewer that reads only the repo enforces the part somebody wrote down. The routine for writing them down by hand is in turning review comments into agent rules.
- Two rulebooks. A review bot's own file (
BUGBOT.md, path instructions, a dashboard) drifts from theAGENTS.mdthe agents read, because different people edit them. - Whether the agent ever saw the rule. A review catches the break after the code is written. It cannot tell you the rule never loaded, which is the usual cause (Codex, Claude Code, Copilot).
Where Harbor fits
Harbor's GitHub App reviews each pull request against the repo's rules files, read from the base branch so a pull request cannot rewrite them: CLAUDE.md, AGENTS.md, Cursor and Copilot rules, and CONTRIBUTING.md. It needs nothing else to start. On a private repo it also checks the rules your team's brain learned from Slack, reviews and agent sessions, which are the same rules Harbor serves to Claude Code, Cursor, Codex and Copilot while they write the code. A public repo is checked against its own files only, so private rules never appear in public comments.
It comments on the lines that break a rule, quotes the rule, and never fails the check, so a broken rule informs a merge and does not block it. It checks rules; it is not a general bug review, and the tools above are better at that. A Free workspace gets 10 reviews. Setup is in the GitHub App tutorial, and what it reads and keeps is in the code review docs.
Questions
Does CodeRabbit read AGENTS.md?
Yes. CodeRabbit reads AGENTS.md, CLAUDE.md, .cursor/rules, copilot-instructions.md and several other agent rule files by default as code guidelines, and .coderabbit.yaml can add more file patterns.
Does GitHub Copilot code review read AGENTS.md?
Yes. Copilot code review reads AGENTS.md along with .github/copilot-instructions.md and .instructions.md files. GitHub's docs say it reads them from the head branch, the pull request itself, not the base branch.
Does Cursor Bugbot follow my Cursor rules?
No. Cursor's docs say project rules in .cursor/rules/*.mdc do not apply to Bugbot runs. Bugbot reads .cursor/BUGBOT.md, and a .md or .mdc file linked from BUGBOT.md counts as a rule.
Can a pull request change the rules it is reviewed against?
With a reviewer that reads instructions from the head branch, yes: a pull request can edit AGENTS.md and be reviewed against the edited file. Anthropic's Code Review reads REVIEW.md from the base branch, and Harbor reads every rules file from the base branch, so a pull request cannot rewrite them.
How do I review pull requests against CLAUDE.md for free?
Run Claude in your own CI with GitHub Actions, set up from Claude Code with /install-github-app, on your runners and API key. Put review criteria in CLAUDE.md, add @AGENTS.md to it if your rules live there, and ask the review to quote the rule behind each comment.