Best Code Review Skills for Claude Code and Agents

For code review in Claude Code, the best options are Brooks Lint for a structured in-session review, Open Code Review for rule-driven reviews from the command line or CI, and Anthropic's Claude Code Action for pull requests on GitHub. All three are free and open source. We read each README and licence on 12 November 2026 and have not run them, so treat this as a guide to what each one claims.

What a review skill does

A plain request such as "review this change" works in any agent. A review skill adds a method. It decides what to read, which checks to apply, how findings are written and sometimes how a score is given. That is useful for two reasons: reviews become more consistent from one run to the next, and findings are easier to act on.

The skills in this guide fall into three kinds, and the kind matters more than the brand.

If you are new to the format, our skills guide explains how skills load, and our shortlist of the best Claude skills covers other jobs. The full directory for this job is code quality.

The shortlist at a glance

Pushed means the date of the last commit to the repository we checked. A recent push does not prove quality, but a long gap is worth a look.

The last column is the one people skip. A review tool that sends your code to a service is a data decision, not only a tooling choice.

In-session skills

Brooks Lint

Brooks Lint grades code against twelve classic engineering books, including The Mythical Man-Month, Refactoring and Clean Architecture. The README describes six decay risks for production code and six for tests. Each finding comes back as symptom, source, consequence and remedy, with a book citation, a severity label and a score out of 100.

It has six commands.

A .brooks-lint.yaml file sets strictness, disables checks and ignores paths such as generated files. Install in Claude Code is /plugin marketplace add hyhmrright/brooks-lint then /plugin install brooks-lint@brooks-lint-marketplace, and an installer script covers more than ten other agents. It is MIT licensed.

Two cautions from the README. A maintainer notice dated 16 October 2026 says only Codex is still tested, so Claude Code and other platforms are not verified now. And the benchmark table in the README is labelled illustrative, so read it as a description of the output format, not a proof of accuracy.

Trail of Bits skills

Trail of Bits publishes skills for security research, vulnerability detection and audit workflows. They give an agent an audit method instead of a general checklist. The licence is CC BY-SA 4.0, a share-alike licence meant for creative works, so check what it means before you copy skill text into your own repository. Findings need confirming by hand before anyone reports them.

Claude Code Workflows

Claude Code Workflows is a repository with three worked workflows: a code review, a security review based on the OWASP Top 10, and a design review that uses the Playwright MCP server to look at front-end changes in a browser. There is no installer, so you copy the pieces you want into a project. It is MIT licensed and suits a solo developer who wants a starting point to adapt.

Command line tools

Open Code Review

Open Code Review comes from Alibaba, where the README says it began as an internal review assistant. It reads your Git diff, sends changed files to a language model you configure, and returns structured comments with line-level positions. Its design choice is that plain code, not the model, decides which files to review, groups related files into units for sub-agents and matches rules to each file. The README presents that as the fix for agents that skip files or report the wrong line.

The commands are simple.

npm install -g @alibaba-group/open-code-review
ocr config provider        # pick a provider and enter a key
ocr review                 # working changes
ocr review --from main --to my-branch
ocr scan --path src        # review whole files, not a diff

It needs Git 2.41 or newer. There are two ways to run the model. In the default mode it uses its own provider settings and API key, not your Claude subscription. In delegate mode, your coding agent does the review with its own model, and OCR only handles file selection and rules. For Claude Code there is a plugin that adds /open-code-review:review and /open-code-review:delegate-review. The licence is Apache 2.0.

The README reports a benchmark on 200 real pull requests with higher precision and F1 than a general agent, and lower recall, which it calls a deliberate trade. That is Alibaba's own result and we have not repeated it. Lower recall means it will miss more real issues, which matters if you plan to rely on it alone.

Skylos

Skylos is a command line tool and pull request gate. skylos . finds dead code such as unused functions, classes and imports, and adding -a checks security bugs, hard-coded secrets, vulnerable dependencies and common mistakes in AI-written code. skylos done checks whether a branch is ready to merge. It runs locally with no account, and there is an MCP server. The README lists 11 languages, with the deepest checks for Python and TypeScript or JavaScript, so look at your own stack first. It is Apache 2.0.

Pull request automation

Claude Code Action

Claude Code Action runs Claude inside GitHub workflows, so it can respond to pull requests and issues. The README's quickstart is to run /install-github-app in Claude Code, which sets up the GitHub app and secrets. You must be a repository admin, and that route is for direct Anthropic API users, with separate guides for Bedrock, Vertex AI and Microsoft Foundry. It is MIT licensed. An agent with write access to a repository deserves tight permissions, so scope the token and require a person to approve what it proposes.

Claude Code Security Reviewer

Anthropic's security reviewer is a separate GitHub Action that reviews pull request changes for security problems. Its README says it is diff aware, comments on the pull request, works with any language and filters low-impact findings that tend to be false positives. You add a workflow that uses anthropics/claude-code-security-review@main and pass an API key as a secret. Inputs let you exclude directories, choose the model and supply your own scan and false-positive instructions. It is MIT licensed.

Two things to weigh. The last push was 11 February 2026, so it has been quiet for about nine months. And the README warns that it is not hardened against prompt injection, so review only trusted pull requests, and consider requiring approval before the workflow runs on pull requests from forks.

CodeRabbit plugin

The CodeRabbit plugin brings CodeRabbit's review into more than 35 coding agents, including Claude Code, Cursor and Gemini CLI. You install the CodeRabbit command line tool, then tell your agent to review your code, and it handles sign-in if needed. In Claude Code the install is /plugin install coderabbit@claude-plugins-official. The plugin repository is MIT licensed. The review itself runs through CodeRabbit's service, so it needs an account and your code is sent there. Check your data policy first.

How to choose

Do not stack all of them. Two tools reviewing the same diff produce two long lists with a lot of overlap, and people stop reading. A workable setup is one in-session review before you push, one automated check on pull requests, and a human who approves the merge.

Limits you should plan for

• False positives. Every tool here can flag fine code. Early on, compare its findings with your own review of the same change, and tune the ignore rules.
• Missed issues. No tool finds everything. Open Code Review's own README says its recall is lower than a general agent's, and the others make no promise either.
• Cost. Reviews are model calls. A whole-repository scan costs more than a diff review. Start with a diff, and check your plan's usage or the API bill after a week.
• Prompt injection. Code under review can contain text meant to steer the reviewing agent. The security reviewer's README says so directly. Do not run an automated agent with write access on pull requests you do not trust.
• Data leaving your machine. Check where each tool sends code. A tool that calls a model provider or a hosted service sends your diff to it.
• Age. A repository that has been quiet for months may not match current agent versions. The security reviewer's last push was in February, while Open Code Review and Brooks Lint were pushed on the day we checked.

A sensible first week

1. Day one. Install one in-session skill and review a change you have already merged. Compare the findings with what your human reviewers said.
2. Day two. Write down which findings were useful, which were noise and which it missed. Adjust ignore rules or strictness.
3. Day three. Try the command line tool on a branch. Time it and note the cost.
4. Day four. Decide whether you want pull request automation. If so, start on a test repository with tight permissions.
5. Day five. Keep one in-session tool and at most one automated check. Remove the rest.

Before you install any of them

• Licence. Read the licence file in the repository. Trail of Bits uses CC BY-SA 4.0, which asks that copies of its text be shared under the same terms. The rest here are MIT or Apache 2.0.
• Where it runs. Note whether the tool runs locally, calls your own model, or calls a vendor's service.
• Keys. Put API keys in your CI secrets, never in a prompt or a committed file.
• Permissions. Give a pull request tool read access first and add write access only if you need it.
• Installers. Clone and read any installer script before you run it.

For other jobs, see best Claude skills. If you want to connect your agent to other tools, our Claude MCP guide covers that.

Our take

Start with an in-session review and an honest comparison. Brooks Lint is the most complete skill we found for that, and its structured output with citations makes findings easy to challenge. Open Code Review is the one to try if you want a repeatable command that also runs in CI, with the trade that it uses its own model setup and key unless you use delegate mode. For GitHub pull requests, Anthropic's own action is the obvious first look, with permissions kept tight.

What none of them replace is a person who understands the change. Use them for the routine problems, so human review time goes to design and intent. We have read these projects closely and not run them, so test any of them on a few of your own past changes before you rely on it.