Blue cover with the title Codex plugin for Claude Code: setup guide, next to a tilted illustration of two developers working at desks in a green garden office. Cover for a guide to installing and using the Codex plugin inside Claude Code.
All articles

Codex plugin for Claude Code: setup, commands, review loop

Install OpenAI's Codex plugin for Claude Code, run /codex:review and /codex:rescue, set the review gate and use both agents on one project. Checked Nov 2026.

The Codex plugin for Claude Code is OpenAI's official plugin that lets you call Codex from inside a Claude Code session. You install it with four slash commands, then run /codex:review for a read-only second opinion on your changes or /codex:rescue to hand a task to Codex. It uses the Codex CLI already on your machine, so the work counts toward your Codex usage limits, not your Claude plan. Everything below was checked against the plugin's README in November 2026.

VibeiDE runs your coding agents in parallel. Watch them as an Office or a Grid.14 days free

What is the Codex plugin for Claude Code?

It is a Claude Code plugin, published by OpenAI in the codex-plugin-cc repository, that adds Codex commands and a Codex subagent to Claude Code. The plugin does not ship its own runtime. It delegates through your local Codex CLI and app server, with the same install, sign-in, repository checkout and environment you would get by running codex yourself.

That design has two practical effects:

  • One checkout, two models. Claude writes or plans, Codex reviews or takes over, and both see the same files.
  • Two separate allowances. A Codex review spends Codex usage. Your Claude limits are only touched by what Claude itself does.

If you are still deciding between the two agents, our Codex vs Claude Code comparison covers sandboxing, approvals and limits side by side.

How do you install the Codex plugin for Claude Code?

Run four commands inside Claude Code: add the marketplace, install the plugin, reload plugins and run the setup check.

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

Requirements, as listed in the README in November 2026:

  • A ChatGPT subscription (the Free plan counts) or an OpenAI API key.
  • Node.js 18.18 or later.
  • The Codex CLI. /codex:setup reports whether Codex is ready and can offer to install it when npm is available. To install it yourself, run npm install -g @openai/codex. Our Codex CLI install guide covers each platform.

If Codex is installed but not signed in, run the login from inside Claude Code with the shell prefix:

!codex login

After the install, the /codex: commands appear in the slash menu and a codex:codex-rescue subagent shows up in /agents. If you have not used plugins or subagents before, see Claude Code subagents.

Which commands does the Codex plugin add?

The plugin adds eight slash commands: two reviews, one task handoff, one session transfer, three job controls and the setup check.

CommandWhat it does
/codex:reviewRead-only Codex review of uncommitted changes, or of a branch with --base <ref>
/codex:adversarial-reviewRead-only review that challenges design choices and accepts focus text
/codex:rescueHands a task to Codex through the codex:codex-rescue subagent
/codex:transferCreates a Codex thread from the current Claude Code session and prints a codex resume command
/codex:statusShows running and recent Codex jobs for this repository
/codex:resultShows the stored output of a finished job, with the Codex session ID when available
/codex:cancelCancels an active background job
/codex:setupChecks install and sign-in, and turns the review gate on or off

/codex:status, /codex:result and /codex:cancel accept an optional job ID, so you can address one job when several are running.

How do you run a Codex review from Claude Code?

Run /codex:review after Claude finishes a change, and Codex reviews the uncommitted diff without editing anything. The README suggests this first run:

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

Three things are worth knowing:

  1. Pick the target. With no flags, the review covers uncommitted changes. Add --base main to review the whole branch against main.
  2. Pick wait or background. --wait keeps the review in the foreground. --background returns control at once, which the README recommends for large reviews.
  3. Use the adversarial review for design questions. /codex:review takes no custom focus text. /codex:adversarial-review does, after the flags:
/codex:adversarial-review --base main challenge whether this was the right caching and retry design
Four step loop: Claude Code drafts a plan, Codex reviews it, the author revises, then you approve and build
Using both agents on one project: one drafts the plan, the other reviews it before any code is written.

The value of a second model is that it is wrong in different ways than the first. A review from the model that wrote the code tends to repeat its own assumptions. Read the findings yourself before you accept them: our guide to reviewing AI generated code lists what to check. For the reviews Claude Code can run itself, see our Claude Code review guide.

What is the review gate, and should you turn it on?

The review gate makes Codex review Claude's answer every time Claude tries to stop, and blocks the stop when Codex finds issues. You control it from the setup command:

/codex:setup --enable-review-gate
/codex:setup --disable-review-gate

It works through a Stop hook, the same mechanism described in our Claude Code hooks guide. When the gate blocks a stop, Claude gets the findings and tries to fix them before finishing.

The README carries a clear warning: the gate can create a long-running loop between Claude and Codex and drain usage quickly, so enable it only when you plan to watch the session. A sensible default is to leave it off and run /codex:review yourself at the points that matter, such as before a commit.

How do you hand a task to Codex with /codex:rescue?

Type /codex:rescue followed by the task, and the codex:codex-rescue subagent passes it to Codex. It suits a bug Claude keeps circling, a second attempt at a fix, or work you want to continue while Claude does something else.

/codex:rescue investigate why the tests started failing
/codex:rescue --resume apply the top fix from the last run
/codex:rescue --background investigate the regression

The flags:

  • --background and --wait work as they do for reviews.
  • --resume continues the latest rescue thread for the repository. --fresh starts a new one. With neither, the plugin may offer to continue the latest thread.
  • --model and --effort choose the Codex model and reasoning effort for that run. Without them, Codex uses its own defaults.

You can also ask Claude in plain language to delegate a task to Codex. Unlike the reviews, a rescue can change files, so treat its result like any other agent edit and read the diff.

To move a whole conversation, /codex:transfer creates a persistent Codex thread from the current Claude Code session and prints a codex resume <session-id> command. Delegated tasks and gate runs can be reopened the same way, with the session ID from /codex:result or /codex:status. Our Codex resume guide explains what a resumed session restores.

Codex terminal with sandbox workspace-write and approval on-request next to Claude Code terminal in default permission mode that asks before edits and commands
Codex draws the line with a sandbox; Claude Code draws it with permission prompts.

Which Codex settings and limits apply?

Your normal Codex configuration applies, because the plugin wraps the global codex binary. Settings come from two files:

  • ~/.codex/config.toml for your user.
  • .codex/config.toml in the project, loaded only when the project is trusted.

For example, to set the reasoning effort for one project:

model_reasoning_effort = "high"

On limits: the README states that plugin usage counts toward your Codex usage limits. As of November 2026, the Codex pricing page describes how each ChatGPT plan meters Codex, and our Codex usage limits guide explains how to check what is left. An API key in your Codex config keeps working too, and is billed per token instead.

What other ways are there to use Codex and Claude Code together?

Yes. There are four kinds of tool for using Codex and Claude Code together on one project: OpenAI's Codex plugin for Claude Code (this post), a shared instructions file that both agents read, running the two CLIs in separate git worktrees, and desktop apps that run both agents on one project, such as VibeiDE, Offrun and Conductor. The plugin is the most direct route. Pick by how much you want the two agents to know about each other.

ApproachHow it worksBest for
Codex pluginCodex runs inside a Claude Code session through slash commandsReviews and handoffs without leaving Claude Code
Shared instructionsAGENTS.md holds the project rules and CLAUDE.md imports itKeeping both agents on the same conventions
Separate worktreesEach CLI runs in its own terminal and its own git worktreeTwo tasks at the same time
Desktop appsAn app such as VibeiDE, Offrun or Conductor runs both agents on one projectMany tasks across several projects

For shared instructions, the Claude Code memory documentation says Claude Code reads CLAUDE.md, not AGENTS.md, and recommends a CLAUDE.md that imports the shared file:

@AGENTS.md

## Claude Code

Use plan mode for changes under `src/billing/`.

For separate sessions, give each agent its own checkout so they never write to the same files. See Codex worktree and Claude Code worktree.

Desktop apps that run both agents on one project, as each tool's own site described it on 19 November 2026:

  • VibeiDE runs Codex and Claude Code in embedded terminals with a task queue per project, and its plan council has one provider draft a plan while the other reviews it critically. The next section has the details.
  • Offrun describes itself as a free app for running coding agents side by side, on Mac, with Windows and Linux in beta. Its site says each agent gets its own git worktree and branch, and one agent can review the other's diff.
  • Conductor says it runs the Claude Code, Codex, Cursor and OpenCode agents from a macOS app, an iOS app and Conductor Cloud. Its site says Conductor is joining Meta and will keep working as normal. See VibeiDE vs Conductor.

One route to skip: older guides start Codex as an MCP server so that Claude Code can call it as a tool. As of November 2026, the Codex MCP server page says the codex mcp-server command and the standalone binary have been removed.

How does VibeiDE use both agents?

VibeiDE is a desktop app for macOS, Windows and Linux that coordinates the Codex, Claude Code and OpenCode CLIs you already installed. Codex and Claude Code run in embedded terminals that show the original CLI interface, and each user signs in to their own provider accounts. Its Plan council has one provider draft a plan, the other review it critically, and the author revise it before you decide to implement. Each project has its own task queue, the branch mode "New worktree per task" gives each task its own Git worktree, and finished tasks wait in Ready for review. There is more for Codex users and Claude Code users.

FAQ

Do I need a separate Codex account for the plugin?

No, if Codex is already signed in on the machine. The plugin uses your local Codex CLI authentication. Otherwise sign in with a ChatGPT account or an API key using codex login.

Does the Codex plugin use my Claude usage?

Codex work started by the plugin counts toward your Codex usage limits. Claude still spends its own usage on reading the results and acting on them, and the review gate adds extra Claude turns each time it blocks a stop.

Can Codex edit my files through the plugin?

The two review commands are read-only. /codex:rescue hands Codex a real task, which can include trying a fix, so review those changes like any other agent edit.

Takeaways

  • Install with /plugin marketplace add openai/codex-plugin-cc, /plugin install codex@openai-codex, /reload-plugins and /codex:setup.
  • Use /codex:review for a read-only second opinion, and /codex:adversarial-review when you want to steer the critique.
  • Use /codex:rescue to hand over a task, and /codex:status, /codex:result and /codex:cancel to manage background jobs.
  • Leave the review gate off unless you are watching: it can loop and drain usage.
  • Plugin work counts toward your Codex limits and follows your Codex config.toml.
  • Skip the old MCP server route: codex mcp-server has been removed.

Share this article

Post on XShare on LinkedIn

Related articles

Necessary cookies support sign-in, security, your language and this choice. Optional categories stay off until accepted.

First-party page views and acquisition measurement, plus Google Analytics 4 (Google Ireland, data may reach the US). Advertising personalization stays off.

With Analytics enabled, retain Google, Microsoft and Reddit ad click identifiers and share download, signup, activation and purchase conversions with the relevant ad provider. No personalized advertising.

Privacy · Cookie inventory