Code review in Claude Code

OpenQodex gives you code review in Claude Code before you push. You ask for a review in Claude Code, or run one command in your terminal, and OpenQodex prints one report. It runs on your own Claude Code login, with no other key, account or server.

What you get

OpenQodex is open source AI code review. One command, openqodex review, works out your change: the commits not yet pushed plus everything uncommitted. It runs the scanners that fit the changed files and keeps only findings on the lines you changed. A scanner is a program that checks code without a model, such as gitleaks or semgrep. Then OpenQodex starts its own reviewer: a separate Claude Code process, on your own login, that reads a frozen copy of the change. The review runs in stages: two finder processes read each changed file with the code around the change, and a fresh verifier process checks every candidate they raise. Only a confirmed defect becomes a finding. A style, naming, docs or test item is shown as an observation, apart from the findings and outside the verdict. Every scanner finding is checked too, and every changed line is given to a finder. Scripts check each answer, and OpenQodex prints one report.

Install

Run this in your own terminal:

npx openqodex init

Or ask Claude Code to run npx -y openqodex@<version> init --yes --agent claude-code (the README gives the current version): the same install, with no question.

init finds Claude Code, Cursor, Codex CLI and Cline on your machine. It prints every file it will write and asks once. In the default user scope, it writes these for Claude Code:

• the skill, in ~/.claude/skills/openqodex/SKILL.md.
• the push gate hook, merged into ~/.claude/settings.json.
• one line in ~/.claude/CLAUDE.md: before any push, review the change with the openqodex skill.
• permission rules in ~/.claude/settings.json, so Claude Code runs review and review --all without asking you.

agents lists every file, the project scope and the uninstall.

Other ways to install:

• The Claude Code plugin. The openqodex/openqodex repository holds a plugin marketplace with an openqodex plugin. The plugin carries the skill and the push gate hook. Its hook calls npx -y openqodex@<version>.
• The skill alone: npx skills add openqodex/openqodex -g. This writes the skill but no hook, launcher or scanner download.

OpenQodex needs Node 22 or newer and git. It runs on macOS and Linux. On Windows, use WSL.

Run a review

Say to Claude Code:

review my change with openqodex

The skill tells Claude Code to run one command, openqodex review, wait for it, and show you the receipt it prints: the verdict, one line per finding and the absolute path of report.html. Then Claude Code asks "Fix all, or tell me which?" and fixes only the findings you name. Claude Code does not review the change itself and starts no subagent. In your own terminal, run ~/.openqodex/bin/openqodex review, the full path init prints.

A review uses your own Claude Code plan. It can take minutes. On the 2026-10-09 benchmark run, a review of a small change took about two minutes with Claude Code. review prints a line every 15 seconds while the reviewer works. openqodex review --all reviews the whole repository instead of one change.

What the reviewer can and cannot do

The reviewer is a fresh Claude Code process (claude -p) with no memory of your session. The reviewer:

• reads a frozen copy of the change in ~/.openqodex/checkouts/, never your folder. Secrets the scanners found are redacted in that copy first.
• has the read, search and list tools, plus Claude Code's WebSearch and WebFetch. It has no shell, no edits, no MCP server and no subagent.
• edits nothing and runs none of the repository's code.
• loads none of your Claude Code settings, hooks, plugins, memory or CLAUDE.md files as agent instructions. Repository files chosen in review.instructions_files are quoted as convention data in the brief. See Convention files.
• gets only the environment variables Claude Code needs to run and find its login. Other tokens in your shell, such as GITHUB_TOKEN or NPM_TOKEN, never reach it.

Claude Code's own permission rules refuse a read outside the frozen copy. OpenQodex also checks every tool call the reviewer makes. It marks the review incomplete when a call names a path outside the copy. The report lists the files the reviewer read.

Web search is on by default. A reviewer that reads private code and untrusted text from the change, and can open web addresses, can be talked into putting that code into a web address. Set this in ~/.openqodex/config.yaml to remove the web tools:

reviewer_web: off

Claude Code sends the review brief and the files the reviewer reads to the model your Claude Code login uses, as any Claude Code session does. The brief is the text OpenQodex gives the reviewer to review from. OpenQodex and the built-in scanners send no code anywhere. security lists every network call.

When Codex is installed and logged in too, a review you run from Claude Code still uses Claude Code. --reviewer codex or reviewer: codex in ~/.openqodex/config.yaml picks Codex.

The push gate

init adds one PreToolUse hook to Claude Code. It matches the Bash tool and runs only for git push commands. The hook looks for the record review wrote for your current work. It does not know what a push sends: for a plain git push it asks whether your current work has a passing review, and for any other push command it says it cannot tell. The git pre-push hook below checks the exact commits a push sends. It never scans, never starts a review, and never approves a push for you.

By default the gate only warns:

• A complete review of this change that passed: the gate says nothing.
• No review of this change: one line asking you to run openqodex review.
• An incomplete review of this change: one line saying so. It never blocks.

The gate blocks only when .openqodex/config.yaml sets review.block_on_severity. Then it denies a push with no review of the change, so Claude Code runs the review and tries again. It also denies a push whose review is blocked, with the counts and the absolute path of report.html, and tells Claude Code to ask you which findings to fix. config explains how to block pushes at a severity.

OPENQODEX_SKIP=1 in the environment lets the push through and says so. It is your switch, not your agent's.

Inside a repository, init also adds a git pre-push hook, so every push from the repository is checked for a review, from an agent or by hand. --hook none leaves it out.

Review a pull request or a branch

To review a pull request or a branch that is not your current work:

openqodex review '#42'
openqodex review <branch>

OpenQodex fetches the pull request or the branch, checks it out in a temporary folder and reviews what it added since it left its base. Your working folder is not touched. When gh is installed, OpenQodex asks it for the pull request's base. A review of a branch or a pull request names its target, so Claude Code asks you before each one.

The same review in a GitHub Action

The OpenQodex GitHub Action runs the review on a pull request when the workflow gives it your Anthropic API key. It runs the full openqodex review: the scanners and a separate Claude Code reviewer, with script checks of its answer. The job summary shows the report, and the findings go to GitHub code scanning. The reviewer's web tools are off there.

Add the key as a repository secret named ANTHROPIC_API_KEY, and set it on the OpenQodex step only. Each review spends your repository's own API credit at Anthropic's prices. Without a key, the Action runs the scanners only (openqodex scan), which is not a review. github-action gives the workflow and every input.

What it does not do yet

• The reviewer needs Claude Code or Codex, installed and logged in. Without either, review prints "Full review unavailable", says what is missing, saves the unchecked scanner findings to a file it names, and names the command with which the agent you are in reviews the change itself (review --agent).
• No review on your own API key without Claude Code or Codex.

Where the report goes

Claude Code shows you the receipt OpenQodex printed: the verdict (passed or blocked), the reviewer's summary, one line per finding with its number, severity, category, title, file and line, and the absolute paths of report.html and report.md.

Open report.html in a browser to read the review. It shows each changed file as a diff with each finding under its line: where it is, the problem, why it matters, the fix and its source. Below are the coverage, the scanners and the blast radius. The page runs no script and loads nothing. Then tell Claude Code which findings to fix, by number, or say "fix all".

The same review is in .openqodex/reviews/<time>-<id>/ in your repository, as report.html, report.md, report.json and report.sarif. .openqodex/.gitignore keeps the reports out of git.

A complete review means every stage ran, every scanner finding was checked and every changed line was put in front of the reviewer. It does not mean nothing was missed. When a review is not complete, review prints "Review incomplete" with what is missing, and exits 2.

Next

• agents: every file init writes for each coding agent.
• config: block pushes at a severity, exclude paths, switch scanners off.
• security: what runs, what is sent where, and where files go.