You don't have to pick one

Run Codex insideClaude Code.

One plans, the other executes, in the same terminal. No tab switching, and no copying code between two chats.

OpenAI ships an official plugin for this. It takes four commands, and it works whether you pay for ChatGPT or not.

Claude Codex

How it is wired

How Claude Code and Codex are wired together Inside one terminal on your machine: Claude Code plans and sends the task to Codex, Codex executes on its own and returns the result, Claude Code reviews it and writes to your files. YOUR MACHINE · ONE TERMINAL CLAUDE CODE reads your files, plans the work sends the task CODEX executes on its own, in the background returns the result CLAUDE CODE reviews it, then writes YOUR FILES

Nothing leaves your machine that would not have left it anyway. Codex runs where you run it, on the files you already have open. The two never talk to each other behind your back: every hand-off is a command you typed.

What you get

Eight commands, inside the session you already use.

Codex stops being another window to open. It becomes something you call when you need it, and it reads the same files you are working on.

/codex:reviewA read-only code review of your current git state. Good before a push.
/codex:adversarial-reviewThe same review, but it argues with your approach instead of checking your syntax.
/codex:rescueHand a bug or an investigation to Codex when you are stuck. It works on its own and reports back.
/codex:transferMove the current session into a Codex thread you can pick up later.
/codex:status · result · cancelAnything you send to the background lives here: what is running, what it found, and how to stop it.
/codex:setupTells you whether the Codex side is actually ready, and turns the stop-time review gate on or off.
Before you start

Two things, and you probably have one already.

  • 1A ChatGPT account. The free tier works. What you run counts against your Codex usage limits, same as if you used Codex directly. An OpenAI API key works too.
  • 2Node.js 18.18 or later. If you followed the setup guide, you already have it. Type node -v in your terminal to check.
Install

Four commands, in this order.

Type them inside Claude Code, one after the other. Each one waits for the previous to finish.

1

Add the marketplace

This tells Claude where to find the plugin. It is the official OpenAI repository, nothing third party.

/plugin marketplace add openai/codex-plugin-cc
2

Install the plugin

This pulls the commands and the Codex subagent into your setup.

/plugin install codex@openai-codex
3

Reload

Without this, the new commands do not show up yet.

/reload-plugins
4

Check it is ready

This one tells you whether Codex is installed and logged in. If Codex is missing, it offers to install it for you.

/codex:setup
If it says Codex is not installed and you would rather do it yourself, run npm install -g @openai/codex in your terminal. If it is installed but not logged in, run codex login.
First run

Give it something real straight away.

Open a project you are working on, then ask Codex to review what you changed. It runs in the background, so you keep working while it reads.

/codex:review --background

When it is done, /codex:result shows you what it found.

The bridge

Hand over the whole conversation, not a summary of it.

This is the part that makes it one setup instead of two tools sitting next to each other. You are twenty exchanges into a job with Claude and you want Codex on it. You re-explain nothing and you paste nothing.

/codex:transfer

It reads the transcript of the session you are actually in, imports it into Codex as an ordinary Codex thread, and hands you back the thread id along with the command that opens it.

codex resume <thread-id>

Run that in a terminal and Codex is standing in the same conversation, holding what the two of you had already worked out. Not a briefing written for it. The history itself. From that point on it behaves like any other Codex thread, so you pick it up again whenever you want.

You hit the ceilingClaude runs out of allowance in the middle of something. The work does not stop, it changes hands and keeps its memory.
You want a real second readCodex arrives with the whole thread, so it argues with the decision you made two hours ago, not with the snippet you pasted.
You are both stuckA different model on the same context tends to see the thing you and Claude walked past four times.

The other direction exists too. /codex:rescue keeps its own thread, so --resume carries on where the last rescue left off instead of starting the explanation again.

Which one for what

They are not interchangeable.

After a few weeks of running both, this is roughly how the work splits for me.

Claude

Building and deciding

Writing the thing, holding the plan across many files, talking through what to do next. It stays in the conversation with you.

Codex

Checking and rescuing

Reviewing before a push, and taking over a bug that has resisted two attempts. A second reader who was not in the room when the decision was made catches things the author cannot.

Not set up yet?

If Claude Code is not running on your machine, start with the setup guide. It goes from nothing installed to a working setup, and you do not need to know how to code.

Open the guide ↗