openai/codex-plugin-cc: run Codex reviews and delegated tasks inside Claude Code

The gap between two coding assistants that do not talk to each other

Claude Code and Codex are separate tools with separate sessions. If you use Claude Code as your main interface but want Codex to look at a change, the default workflow is to leave the session, open Codex, and reconstruct the context by hand. This plugin exists to remove that context switch. The README states the target user plainly: Claude Code users who want to start using Codex from the workflow they already have. It is not a general Codex client and not a plugin for other editors. Everything it does is expressed as slash commands inside Claude Code, plus one subagent named codex:codex-rescue that appears in /agents. The requirements are narrow and worth reading before anything else: a ChatGPT subscription, including Free, or an OpenAI API key, and Node.js 18.18 or later. The README also notes that usage contributes to your Codex usage limits, so the plugin does not give you a second free quota. It spends the quota you already have.

Reviews, rescue tasks and background jobs: how the commands divide up

The command surface splits into three jobs. Review commands are read-only. /codex:review runs a normal Codex review on your current uncommitted changes or on your branch against a base ref, and the README says it gives the same quality of review as running /review inside Codex directly. It accepts --base <ref>, --wait and --background, and it is explicitly not steerable: you cannot add focus text. /codex:adversarial-review uses the same target selection but is steerable, taking extra focus text after the flags so you can challenge a specific decision. The README suggests using it for design choices, hidden assumptions and risk areas such as auth, data loss, rollback, race conditions and reliability. Delegation is the second job. /codex:rescue hands a task to Codex through the codex:codex-rescue subagent, supporting --background, --wait, --resume and --fresh, and you can also simply write a request like asking Codex to redesign the database connection. If you pass no --model or --effort, Codex picks its own defaults; the string spark maps to gpt-5.3-codex-spark. Job management is the third: /codex:status shows running and recent jobs for the current repository, /codex:result returns the stored output and, when available, the Codex session ID for codex resume, and /codex:cancel stops an active background job. All three accept an optional task ID such as task-abc123. One design detail stands out: /codex:transfer creates a persistent Codex thread from the current Claude Code session and prints a codex resume command, using Codex's external-agent session importer. The README notes the source must live under ~/.claude/projects, and that older Codex versions without session import must be upgraded first.

Installing the plugin and running a first background review

Installation happens inside Claude Code through the plugin marketplace, not through npm. The README gives the marketplace add first, then the install, then a reload. The plugin name in the install command is codex@openai-codex, which is not the same string as the repository name, so copy it exactly.

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

After reloading, run the setup command. According to the README, /codex:setup reports whether Codex is ready, and if Codex is missing and npm is available it can offer to install Codex for you.

/codex:setup

If you would rather install Codex yourself, the README gives the global npm install, and a login step for an installed but unauthenticated Codex.

npm install -g @openai/codex
!codex login

The exclamation mark prefix runs the command in your Claude Code session rather than through the plugin. Once setup succeeds, the README says you should see the slash commands and the codex:codex-rescue subagent in /agents. A sensible first run is a background review followed by status and result, which the README lists as one simple first run.

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

The README recommends the background path for reviews, noting that multi-file changes might take a while. If you want a branch comparison instead, /codex:review --base main reviews your branch against main.

The review gate and the setup flags the README only partly documents

The README mentions an optional review gate managed through /codex:setup --enable-review-gate and --disable-review-gate, and then the text is truncated. That matters. A gate is the kind of feature that changes when and whether work proceeds, and the available documentation shows the two flags without describing what triggers the gate, what happens when it fails, or how it interacts with background jobs. If you plan to enable it, treat the flag as the starting point of your own investigation rather than a complete specification. The same caution applies to the plugin's lifecycle. There is an install path and a reload command, but the README does not document rollback or uninstall, so a team that needs a documented reversal path has to work that out independently. On the positive side, the review commands are read-only, which the README states twice, and that limits the blast radius of the review half of the plugin. Delegation is different: /codex:rescue is described as investigating, fixing and applying patches, so it writes to your repository. The README's own note recommends forcing long rescue tasks into the background.

Where this plugin is the wrong tool

The plugin assumes you already have Claude Code and are willing to add Codex alongside it. If your goal is to reduce the number of assistants in your workflow, this does the opposite: it makes two models part of the same session, and the README is explicit that usage contributes to your Codex usage limits. Cost and quota are shared with your existing Codex use, not additive. The second limit is environmental. There is no standalone CLI, no Docker image and no self-hosted deployment in the README; installation is a Claude Code marketplace operation, and the plugin is marked private in package.json, so it is not published as a general npm package for other hosts. The third limit is scope. /codex:review is not steerable by design, and the README directs you to the adversarial variant when you want to challenge a specific decision. If your review process depends on directing attention to one file or one risk, the plain review command will not do it. Finally, /codex:transfer depends on Codex's session importer and requires the source transcript to be under ~/.claude/projects, so it is not a general export mechanism.

How it compares with running Codex directly or using a single assistant

The obvious alternative is to skip the plugin and run Codex on its own, either in its terminal UI or through the Codex App. That path gives you the full Codex surface without a Claude Code session in between, and the README's own framing supports it: the plugin is for people who want Codex from the workflow they already have. If your workflow is not Claude Code, the plugin adds a layer and buys you nothing. The second alternative is to stay entirely inside Claude Code and use its own /review. The README claims /codex:review matches the quality of /review inside Codex, which is a claim about equivalence to Codex, not a claim that it beats Claude Code's review. The difference in approach is the model doing the reviewing and the fact that Codex runs as a separate job you can background, poll with /codex:status and cancel. The third comparison is /codex:transfer against manually reconstructing context. The plugin reuses Codex's external-agent session importer and follows the same conversion rules as importing Claude history in the Codex App, producing visible turns you can continue in the App or TUI. Doing that by hand is possible but not equivalent.

Maintenance, versioning and what Apache-2.0 means here

The repository is not archived, and the last push was on 2026-07-08, which is when v1.0.6 was released. Before that, v1.0.5 landed on 2026-06-23 and v1.0.4 on 2026-04-18. The release spacing suggests maintenance activity rather than a frozen snapshot, but the README does not publish a support policy, a deprecation schedule or a compatibility matrix for Codex versions. The one compatibility statement present is that older Codex versions without session import must be upgraded before /codex:transfer will work. The package.json sets engines.node to >=18.18.0, matching the README requirement, and the build script depends on a prebuild step that runs codex app-server generate-ts to produce TypeScript types, which means building from source requires a working Codex binary. The project is licensed Apache-2.0 and ships a NOTICE file alongside the LICENSE. Apache-2.0 permits commercial use and modification and includes an explicit patent grant, with the usual obligations around preserving notices and stating changes. Because the plugin drives Codex, which is a separate OpenAI product with its own terms and usage limits, the licence on this repository does not govern what happens when a task is sent to Codex. That distinction is worth confirming with whoever owns your OpenAI account, and it is not legal advice.

Editorial conclusion

Adopt it if you already work in Claude Code and want Codex reviews or delegated fixes without leaving that session, and you accept that Codex usage counts against your Codex limits. Skip it if you want a single assistant, if you need a plugin that runs without a ChatGPT subscription or OpenAI API key, or if you need a documented uninstall and rollback path, since the README does not cover either. Before relying on it, run /codex:setup and confirm it reports Codex as ready and authenticated, then check whether the review gate is enabled or disabled for your repository.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. Release notes