Tutorial: Install Codex into Claude Code and let two programming agents review each other教程：把 Codex 装进 Claude Code，让两个编程智能体互相复核

Open Claude Code and execute three commands in sequence: /plugin marketplace add openai/codex-plugin-cc to add the OpenAI official market, /plugin install codex@openai-codex to install the plug-in, and /reload-plugins to reload the plug-in for the new command to take effect.

Also confirm the prerequisites: Node.js 18.18 and above; ChatGPT subscription (including free files) or OpenAI API Key. The plug-in body is only about 118 KB. What really does the work is the Codex body, and it consumes your Codex usage.

Execute /codex:setup. It will clearly tell you the readiness status: if Codex is not installed and npm is available on your local machine, you can directly let it install it on your behalf; if it is installed but not logged in, use !codex login to complete the login (the exclamation mark prefix indicates that local commands are executed in Claude Code).

Don’t skip this step: Send a task if the self-check fails. This usually manifests as a silent failure of the background task and only returns a status code, which is very time-consuming to troubleshoot. You can install it manually: npm install -g @openai/codex.

After modifying a piece of code, execute /codex:review to get the read-only review conclusion; if you want to be more aggressive, use /codex:adversarial-review, which will specifically refute this implementation with the questioning angle you specified.

The value lies in the word "independent": the review is completed independently by the Codex model, without looking at Claude's context and self-report, so it is not easy to overlook similar problems because of the preference of the same model. The conclusion is returned directly to the Claude Code session and can be modified in place.

If you encounter problems with Claude repeatedly making incorrect changes, use /codex:transfer to transfer the entire current session to Codex; or use /codex:rescue to dump a description and let it start working independently.

Long-term tasks will go into the background: /codex:status to check the progress, /codex:result to get the results, and /codex:cancel to cancel. It is recommended to try it on the temporary branch first to confirm that the output quality and quota consumption are within the acceptable range, and then put it into the daily process.