The code-review plugin has several specialized agents review an open PR in parallel and filters out low-confidence findings. It needs the gh command and a GitHub repository. It is separate from the built-in /code-review command.
How it works
According to the plugin README, it first skips closed PRs, draft PRs, trivial or automated PRs, and PRs it has already reviewed. Then it gathers the repository's CLAUDE.md guideline files and summarizes the PR changes.
Next it launches four agents in parallel. Agents 1 and 2 both check CLAUDE.md compliance (a redundant check), agent 3 looks for obvious bugs in the changes, and agent 4 uses git blame and history to find problems that depend on context.
Each issue gets a confidence score from 0 to 100, and anything below 80 is dropped. Pre-existing issues, nitpicks a linter would catch, and code with lint-ignore comments are filtered out as false positives.
Install and use
You need a git repository with a GitHub remote, and the GitHub CLI installed and logged in (gh auth login). A CLAUDE.md is optional, but having one improves the guideline checks.
You can change the threshold of 80 and the review criteria by editing the plugin's commands/code-review.md. The README also explains how to add agents for areas like security or performance.
/plugin install code-review@claude-plugins-official /code-review # print results in the terminal /code-review --comment # post results as a PR comment
How it differs from the built-in /code-review
Claude Code also has a built-in /code-review command (alias /review) that reviews the current diff or a PR number at a chosen effort level. This plugin is a separate setup that happens to share the name. I could not confirm which one runs when you mix both in the same environment, so check the output format the first time you use it to tell them apart.
When it is not a good fit
- Checking local changes that have no PR yet. The plugin targets open PRs.
- Replacing human review. The README also says to use the results as a starting point for human review.
- Very large PRs. They take a long time, and the README recommends splitting the PR.
- Non-GitHub repositories (GitLab and so on). The requirements are a GitHub remote and gh.
Real examples
You opened your PR and want a look-over before merging
/code-review
The agents review the open PR for the current branch and list only issues scoring 80 or higher in the terminal. Each item has a file location link, and if it is a CLAUDE.md violation, the rule is named as well. If there are no problems, it may finish with no findings.
You want the review results left where the team can see them
/code-review --comment
It runs the same review and posts the results as a PR comment. The comment is posted from your GitHub account, so it is safer to check the results in the terminal first and post afterward.
Common mistakes and when not to use it
- If you are not logged in to gh, it cannot fetch PR information. Check with gh auth status first.
- Draft PRs and already-reviewed PRs are skipped. Tell the difference between "no findings" and "skipped."
- The confidence filter can drop minor but real issues. Lowering the threshold increases false positives.
- --comment actually posts on the PR. If you post by mistake, you have to delete the comment yourself.
FAQ
The review takes a long time.
Several agents run, so larger PRs take longer. The README recommends splitting PRs into smaller ones.
Can I make it review against my team's rules?
Write the rules in the repository's CLAUDE.md. Agents 1 and 2 check compliance with them.
Checked against the official docs on 2026-10-30. Sources: claude-plugins-official: code-review README, Claude Code docs: Install and manage plugins. How we write