OpenAI just shipped codex-plugin-cc, an official plugin that lets you run Codex from inside Claude Code to review changes, delegate long tasks, or get an adversarial second opinion, and it raced past 26,000 stars within days. That a rival lab is publishing a first-party plugin for Anthropic's coding tool is the story; the setup is genuinely four slash commands. This tutorial gets Codex answering from your Claude Code session in a couple of minutes.

  • It is an official OpenAI plugin that delegates to your local Codex CLI, using the same sign-in and repo checkout.
  • Install is four slash commands inside Claude Code: add marketplace, install, reload, run setup.
  • You need a ChatGPT subscription (Free tier works) or an OpenAI API key, plus Node.js 18.18 or later.
  • Commands like /codex:review and /codex:rescue add a second model's eyes to your workflow.
Single-model versus dual-model codingWithout the plugin, Claude Code reasons with one model. With codex-plugin-cc, you delegate reviews and rescues to Codex through your local CLI, so two frontier models cross-check the same change. ONE MODELClaude Code onlyOne perspectiveNo cross-checkBlind spots persist CODEX PLUGINClaude + CodexSecond opinionAdversarial reviewDelegate long tasks Two frontier models reviewing the same diff, from one prompt genztech.blog
Fig 2 Without the plugin, Claude Code reasons with one model. With codex-plugin-cc, you delegate reviews and rescues to Codex through your local CLI, so two frontier models cross-check the same change.

What is codex-plugin-cc and why does it matter?

The plugin bridges Claude Code and OpenAI's Codex. It does not bundle a separate runtime: it delegates through the Codex CLI already installed on your machine, so it reuses your existing authentication, your repository checkout, and your local environment. Practically, that means you can ask Claude Code to write a change and then hand the diff to Codex for an independent review, or fire off a background task and keep working. It matters because the interesting frontier of AI coding in 2026 is not one model winning, it is orchestrating several. An official cross-vendor plugin makes that multi-model workflow a supported path instead of a hack.

RelatedStrix Setup: Run an AI Penetration Tester on Your Code

How do you install it in Claude Code?

Run these inside a Claude Code session. The setup command tells you whether Codex is ready and can install it for you if npm is available:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

If you would rather install and authenticate Codex yourself:

npm install -g @openai/codex
# then, from the shell inside Claude Code
!codex login

Once ready, you get the Codex slash commands and a rescue subagent. A good first run:

/codex:review --background
/codex:status
/codex:result

Which commands are worth knowing?

/codex:review reviews your working changes, optionally against a base branch with --base main or in the background. /codex:adversarial-review goes further, actively challenging whether your approach was right, which is where a different model earns its keep. /codex:rescue investigates and fixes failing tests, and accepts a model and effort, for example --model gpt-5.4-mini --effort medium. /codex:status and /codex:result track and retrieve background jobs. The pattern that pays off is delegating a slow, well-scoped task to Codex in the background while Claude Code keeps you moving on the next thing.

What are the gotchas?

Three. Codex usage counts against your OpenAI or ChatGPT limits, so heavy background runs draw down the same quota as using Codex directly. You need Node 18.18 or newer for the CLI. And because the plugin rides on your local Codex install, its behavior and any custom base-URL config follow whatever Codex is already set to, which is convenient but means a broken Codex setup shows up as a broken plugin. Fix Codex first, then the plugin just works.

RelatedFirecrawl Setup: Turn Any Website Into LLM-Ready Data

When is a second model actually worth it?

Not every task benefits from two models, and knowing when it does is the difference between a smarter workflow and wasted tokens. The delegation pays off most on three kinds of work. Reviews are the obvious one: a model that did not write the code has no attachment to its choices and spots things the author's model rationalized away. Long, well-scoped background tasks are the second: hand Codex a slow refactor or a test-suite investigation and keep working in Claude Code rather than blocking on it. The third is disagreement on approach, where /codex:adversarial-review earns its name by arguing against your design instead of polishing it. Where a second model adds little is on small, unambiguous edits, there, the round-trip is pure overhead. Treat Codex as a reviewer and a background worker, not a second author for every line, and the usage stays proportional to the value.

What to watch · 2026
  • Cross-vendor plugins. Whether more labs ship first-party plugins for rival agents.
  • Background orchestration. How well long delegated tasks hold up in real projects.
  • Usage limits. Whether dual-model workflows strain free-tier quotas in practice.

Our take

The most telling detail here is who published it. OpenAI shipping an official Codex plugin for Anthropic's Claude Code is a quiet admission that developers already run multiple models and want them to cooperate rather than compete for one chat window. The adversarial review command is the standout, because the real value of a second model is disagreement, not consensus. If you already pay for both ecosystems, this is a near-zero-effort upgrade to your review loop. If you only use one, it is a low-stakes way to see whether a second opinion changes your mind.

Primary sources

Original analysis by GenZTech. Tool documentation: openai/codex-plugin-cc on GitHub.