Codex Plugin for Claude Code

Why it's ranked here

This is a well-scoped wrapper, not a reimplementation: it never runs its own review logic, it forwards work to a Codex installation the developer already authenticates and pays for, and its own tests pin exact wording and behavior for every command, including a fix for a routing bug that had previously caused one command to call itself in a loop. That discipline is what puts it here, but the plugin does nothing without a Codex account already logged in on the machine, which limits it to users already inside OpenAI's ecosystem.

What's good

The read-only review commands are backed by tests asserting they never touch files, only report findings, which is a real guarantee rather than a documentation claim. The tool also encodes its own scar tissue: a regression test locks in the fix for a bug where one command recursively triggered itself through ambiguous routing. Commands accept plain flags for background execution, branch comparisons, and resuming a previous session, so a developer can fire off a review or a bug investigation and check back later instead of waiting on it.

Tradeoffs

Everything routes through a Codex installation the user must already have working, so the plugin adds nothing on its own if that access, subscription, or login is missing. An optional feature that automatically re-runs a review before Claude can finish responding is flagged by the maintainers themselves as able to spiral into a long back-and-forth that burns through usage limits fast, and it is off by default for that reason. The project's own changelog has not kept pace with its release history, listing only the first version.

How to use it well

Turn to this when a task is stuck, tedious, or better handled by a second model with fresh eyes: a flaky test, a bug that resists a first fix, or a pre-ship review you want to run in the background while you keep working. The adversarial mode is worth reaching for specifically before high-stakes changes, like anything touching authentication or data loss, where a second opinion on the design pays off. Leave the automatic re-review loop turned off unless someone is actively watching the session, since it can keep running unattended.

Observed

License

Apache-2.0 license, declared in package.json and included in full as LICENSE.

Runtime requirement

Requires Node.js >=18.18.0; package is an ES module (type: module).

Distribution

Distributed as a Claude Code plugin via a marketplace manifest (.claude-plugin/marketplace.json), not as a standalone CLI package.

Interface

Interface is a set of slash commands plus one registered subagent; no HTTP API or standalone binary is defined in the manifest.

Test suite

Test suite runs on Node's built-in test runner (tests/commands.test.mjs, invoked via node --test) rather than a third-party framework.

Changelog

Changelog (plugins/codex/CHANGELOG.md) contains a single entry documenting only the initial release.