Codex Plugin for Claude Code: /codex:rescue and /codex:review
OpenAI’s Claude Code plugin lets Claude hand tasks to Codex. Here’s how to install it, what its commands do and why to leave the review gate off.
Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.
👁 113 reads
Since mid-October, I’ve had a regular work routine: Claude Code plans a task, I copy its prompt into Codex in another window, wait for the result and paste it back. It worked, but I was acting as the courier. OpenAI has released an official plugin that removes that step: Claude Code can now send a task to Codex, wait for the result and bring it back into the conversation.
I installed it on November 3. Here’s what it includes, how it uses tokens and your usage limit, what to leave switched off, and what went wrong on my first attempt.
Codex plugin for Claude Code at a glance
- Install:
/plugin marketplace add openai/codex-plugin-cc, then/plugin install codex@openai-codex - Requirements: Node.js 18.18 or later and a ChatGPT account, including a free account, or an OpenAI API key
- Main commands:
/codex:rescuedelegates a task;/codex:reviewreviews code - Cost: Plugin tasks count toward your Codex plan limit
- Leave off: The review gate that runs after every response, unless you’re monitoring the session
Install the plugin with four commands
Enter these commands in Claude Code, in order:
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setupThe last command checks whether Codex is installed and signed in. If it isn’t, the command offers to install it with npm install -g @openai/codex. I installed the plugin from the terminal using claude plugin marketplace add and claude plugin install, because I work in the desktop app. The result is the same. If you’re new to Codex, start with this guide to installing Codex CLI on Windows.
It installed version 1.0.6. The command claude plugin details codex@openai-codex shows what’s included: 11 skills, one subagent called codex-rescue and three hooks - scripts that run when a session starts, when it ends and after each response. The plugin adds about 451 tokens to every Claude Code session as a fixed overhead. These tokens describe the commands so Claude knows when to use them. The rest loads only when needed.
Eight commands, three for everyday work
/codex:rescue is the plugin’s core command. It sends Codex a task, such as investigating a bug, fixing it or finishing a piece of work. You can specify a model (--model) and reasoning effort (--effort), run a task in the background (--background) or return to an earlier Codex thread (--resume). The result appears in the Claude conversation without modification, so you can see exactly what Codex wrote.
/codex:review reviews changes in a Git repository, either the current changes or an entire branch against another branch (--base main). It’s read-only and doesn’t make fixes. Its more hard-nosed counterpart, /codex:adversarial-review, looks beyond typos to challenge the underlying idea: whether the chosen approach makes sense, what assumptions it depends on and where it might fail. It’s the most interesting command to me because a model from another company doesn’t share the first model’s habits.
The other commands are for supporting tasks. /codex:status shows tasks in progress, /codex:result retrieves the result of a completed task, and /codex:cancel stops a background task. /codex:transfer moves the current Claude session into a new Codex thread and gives you the codex resume command to open it in Codex.
The review gate can burn through your usage limit
The plugin includes an option that sounds sensible but can be risky in practice. The review gate (/codex:setup --enable-review-gate) makes Codex review every Claude response, then asks Claude to revise anything Codex flags. The hook can run for up to 15 minutes. The developers themselves warn in the documentation that the gate can get Claude and Codex stuck in a long loop and quickly use up your limits.
The gate is off by default. I checked the plugin code, where the setting starts at false, and the /codex:setup report, which shows reviewGateEnabled: false. I left it off. I ask Codex to review manually when I finish a substantial piece of work, not after every sentence.
Every plugin task runs the local Codex CLI on your account, so it uses the same weekly limit as work in the Codex app. I explain how to make that limit go further in Codex settings to help you avoid burning through your limit.
My first task took 25 seconds
For my first test, I gave Codex a useful task through the plugin: review a script for syncing skills that I discuss in How to move Claude Code work to Codex, and list no more than five real bugs. The /codex:setup report was all green: Node.js, npm, Codex CLI 0.162.1 and an active ChatGPT sign-in.
After 25 seconds, Codex replied that it couldn’t read the file. Two attempts ended with the error setup refresh had errors. The plugin wasn’t at fault: the problem was Codex’s Windows sandbox, which isolates commands from the rest of the system. codex doctor showed the issue directly: sandbox provisioning failed. I’d seen the same error before when running Codex from the terminal in read-only mode. On Windows, it’s worth running codex doctor and fixing the sandbox before you first use the plugin. The /codex:setup report doesn’t check it.
Despite that false start, I’m keeping the plugin. Claude is good at planning and maintaining context across long projects, while my Codex Pro plan has its own limit, which I’m already paying for each month. The plugin connects the two tools without me having to shuttle between them.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- OpenAI, codex-plugin-cc (README and version 1.0.6 plugin code): https://github.com/openai/codex-plugin-cc
- My installation and first-task test, using Claude Code 2.1.296 and Codex CLI 0.162.1, November 3, 2026
