Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What is AI code review?

AI code review is a model reading a pull request's diff and the code around it, then leaving comments where it finds bugs, security problems or broken edge cases. Codex, GitHub Copilot, Cursor's Bugbot and Claude Code all offer one. The reviewer comments; a person still decides whether to merge.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

A pull request goes to a reviewer agent, which posts inline comments; a person reads them and decides whether the change merges.runs in the vendor'scloudthe merge decisionstays with a personthe authorthe pull requestreviewer agentCodex, Copilot, Claudeyou, the reviewer1. opens or pushes2. diff and nearby code3. inline comments4. findings, by severity5. merge or send back
  1. A pull request opens or gets a push, or someone comments a trigger such as @codex review, @claude review or bugbot run. Claude's needs write access.

  2. The reviewer reads the diff and the surrounding code. Most also read guidance files in the repository, such as AGENTS.md, REVIEW.md or .cursor/BUGBOT.md.

  3. It posts inline comments on the lines concerned, usually with a severity. Codex limits GitHub comments to P0 and P1; Copilot labels High, Medium or Low.

  4. Some can act on a finding: Copilot's Fix with Copilot and Bugbot's Autofix hand it to a cloud agent that pushes a fix to a branch.

  5. A person reads the findings and merges. Claude Code Review never approves or blocks; Copilot's approving reviews are off by default.

Source: Claude Code docs: Code Review

Example: one pull request, four reviewers

Say a teammate opens a pull request that changes how sessions expire. With Codex, you comment @codex review, or turn on Automatic review so it runs on every PR; it posts a standard GitHub review and, on GitHub, flags only P0 and P1 issues so the thread stays short. With Copilot, you add Copilot as a reviewer (or gh pr edit --add-reviewer @copilot) and get comments labelled High, Medium or Low, often with a suggested change you can apply in a click. Cursor's Bugbot runs on each PR update, or when someone comments bugbot run, and every finding has a Fix in Cursor link. Claude Code gives you two routes: /code-review in your own session, which reviews your branch before you even push, and the managed Code Review service, where several agents read the PR in parallel, a verification step drops false positives, and the rest arrive as inline comments ranked by severity.

Four tools, one shape: read the change, comment on specific lines, leave the decision to a person.

Why it matters

A human reviewer reads a diff with the question "would I have written this?" in mind. An agent reads it with "what breaks?" and has time to open the files around the change, which is where most real bugs hide. That makes it a good first pass: the obvious null check, the edge case nobody tested, the secret pasted into a config. Then the human review gets to spend its attention on design and intent.

The vendors are candid about the limit. Anthropic's docs say a review with no findings doesn't mean the code is error-free, and OpenAI says review rules don't replace tests or required approvals. Treat the reviewer as a colleague who is fast and tireless but can be wrong.

AI code review versus running CI

CI executes your code and tells you whether the tests pass. A reviewer agent reads your code and tells you what it thinks is wrong. Claude Code Review states outright that it works from the diff and surrounding code without building or running anything. So the two find different things: CI catches the regression you had a test for, the reviewer catches the one you didn't. You want both, and neither should be allowed to merge on its own.

Common mistakes

  • Letting the PR edit the reviewer's instructions. Claude Code Review reads REVIEW.md from the base branch and never from the PR's own changes, which closes an obvious loophole. Copilot reads its instruction files from the head branch. Know which your tool does before you trust it on untrusted PRs.
  • Replying to the bot. Copilot's docs say it can't see replies to its comments and won't respond. Fix the code or dismiss the finding instead.
  • Turning the effort up everywhere. Low effort means fewer, more confident findings; high means more coverage and more noise. Claude Code's /code-review and Bugbot both expose this dial. Match it to how critical the repository is.
  • Reviewing without rules. Every tool reads a guidance file. Without one it reviews against generic standards, and you spend the first month teaching it your conventions one dismissal at a time.

FAQ

Does AI code review run my tests?
Usually not. Claude Code Review states that it works from reading the diff and the surrounding code without building or running it, and OpenAI says Codex's review rules don't replace tests or branch protections. Keep CI for what executes; use the reviewer for what reads.
Can AI code review approve a pull request?
Mostly no, by design. Claude Code Review never approves or blocks. Copilot posts a comment-only review by default, and its approvals are a preview feature an admin has to turn on. The point is to inform the human reviewer, not replace the required approval.
How do I tell an AI code review tool what to look for?
With a file in the repository. Codex reads a Code Review Rules section in AGENTS.md, Copilot reads .github/copilot-instructions.md and AGENTS.md, Bugbot reads .cursor/BUGBOT.md, and Claude Code Review reads REVIEW.md. Say what matters in this codebase and what to ignore.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.