Code review that catches what humans skim past.

Code Review reads every pull request and flags the two things reviewers rarely check: comments that no longer match the edited code, and docs the diff just made stale.

Alpha · GitHub App · Free to install

Install the GitHub App

Two checks, zero failing checks.

Every finding is advisory: the bot comments, and never fails a status check or requests changes. It leaves precise, inline evidence and gets out of the way.

Stale comments

When a PR edits code, the comment directly above it often stops being true. The bot re-reads each edited region and flags divergence as an inline review comment, with a suggested rewrite.

Documentation drift

The whole diff is checked against your README, CLAUDE.md, AGENTS.md, and other agent docs. A renamed flag, a changed default, a removed command: flagged in the PR review body.

Honest coverage

If part of the diff could not be analyzed — an oversized file, a capped listing — the result says so. A clean report never silently means “we did not look”.

Findings live where coding agents read.

Results are posted as a real pull request review — inline comments anchored to the changed lines, plus one summary comment that updates in place on every push. Claude Code, Codex, and other agents pick the findings up in their normal review pass; nothing lives in a dashboard they cannot see.

Reviews are pinned to the commit they analyzed. A push mid-review supersedes the run instead of overwriting newer results.

Safe on any repository.

The bot reads diffs and file contents through the GitHub API — it never clones or executes your code, so fork pull requests are reviewed too.

Never runs your code

Analysis is read-only over the GitHub API. No checkout, no install scripts, no build of the reviewed repository.

Untrusted text stays inert

Everything derived from the PR or the model is rendered as literal text before posting — no live mentions, no markdown injection.

Your code stays yours

Repository content — file paths included — is redacted from the operator's error telemetry and logs. Your code goes to the review model and nowhere else.