You know when to call Codex, which command to type, and what you send it.
What you need
- The plugin installed: guide No. 08 Plug Codex into Claude Code, fifteen minutes.
- A ChatGPT account, even a free one, or an OpenAI API key (OpenAI help (new tab), checked on October 6, 2026).
- A folder tracked by Git: reviews cover your changes.
Part 01 / 06
Why two models
Nobody grades their own homework well. A model doesn't either. The second one is not smarter: it wasn't there when the decision was made, so it sees what the author no longer sees.
OpenAI's official plugin, openai/codex-plugin-cc (new tab), puts Codex inside the Claude Code session you already have open. It is not a second piece of software: it calls the Codex installed on your machine, with your account and your settings.
| Who | What I hand over |
|---|---|
| Claude | Building and deciding: writing, keeping the plan across several files, talking through what's next with me. |
| Codex | Reviewing and troubleshooting: the review before a push, the bug that survived two attempts. |
Part 02 / 06
The commands, and when I type them
Eight commands come with the plugin. Here is when I reach for each one.
| Command | When |
|---|---|
/codex:review | Before a push. A read-only review of your changes. --base main compares against a branch, --background runs it in the background. |
/codex:adversarial-review | When the doubt is about the approach, not the syntax. Unlike the plain review, you can tell it where to dig. Read-only too. |
/codex:rescue | When a bug resists. Codex investigates and tries a fix. --resume picks up the last thread, --fresh starts from zero. |
/codex:transfer | When Claude's quota runs out mid-task. The whole conversation moves into a Codex thread. |
/codex:status, result, cancel | To follow what is running in the background, read the result, or stop it. |
/codex:setup | To check that Codex is ready, and to turn the review before stopping on or off. |
The review that pushes back is the one I use most. I give it the question that is nagging me:
/codex:adversarial-review --background look for what breaks if the file is empty
Then /codex:result when it has finished. I keep working while it reads.
Part 03 / 06
The handover when the quota runs out
Claude is out of quota in the middle of a task? The work doesn't stop. /codex:transfer reads the current conversation, turns it into a Codex thread, and gives you the command to resume it.
/codex:transfer
codex resume <session-id>
Codex arrives with the whole history, not a summary you wrote for it. It can discuss the decision from two hours ago, not just the last excerpt.
Plan BIf you get stuck
No thread created? Update Codex: older versions can't import a Claude session.
Part 04 / 06
The review before stopping
One plugin option has Codex review every Claude answer before it stops. If it finds a problem, Claude has to fix it first.
/codex:setup --enable-review-gate
Part 05 / 06
The three-round loop
For longer jobs, I wrote a homemade skill, /route. The planner is never the executor:
- Step 5.1: Claude writes the plan.
- Step 5.2: Codex codes.
- Step 5.3: Claude reviews the result, with
git diffand/codex:review. - Step 5.4: Claude fixes what is off, then we go around for another round.
Part 06 / 06
What goes to OpenAI
Codex runs on your machine, but what it reads is sent to OpenAI to be processed, just as Claude Code sends its reads to Anthropic. Your code therefore goes through two companies instead of one. On a client project, ask yourself the question before typing the command.
Claude can also call Codex by itself: the plugin adds a helper, codex:codex-rescue, visible with /agents. And everything you launch counts against your Codex quota, as if you were using Codex directly.