# OpenAI ships Codex into Claude Code — two commands, or four? > Source: > Published: 2026-07-22 03:53:41+00:00 A rare thing happened on March 30, 2026: OpenAI shipped first-party tooling straight into a competitor's terminal. The result is `codex-plugin-cc` , and it changes how a two-agent workflow fits in one window. `codex-plugin-cc` is an official OpenAI plugin that runs the Codex coding agent from inside Anthropic's Claude Code CLI. It lives at [openai/codex-plugin-cc](https://github.com/openai/codex-plugin-cc), is Apache-2.0 licensed, and is effectively pure JavaScript . The notable part is the direction of travel — a frontier-model vendor packaging tooling for a rival's ecosystem. It is not a new runtime. As OpenAI puts it in its developer-forum announcement: "It is not a separate runtime. It is Codex, just invoked from inside Claude Code." — OpenAI (source: [OpenAI Developer Community]) The plugin delegates to the Codex CLI and app server already installed on your machine, reusing your existing auth and configuration . Maintenance is active: `v1.0.0` landed March 30, 2026, and the current `v1.0.6` shipped July 8, 2026 — six patch releases in four months . This is the same Codex agent that hit more than 5 million weekly users by June 2026 , surfaced in a different terminal. The prerequisites are modest but specific. You need Node.js 18.18.0 or later — this is a hard engine constraint declared in the plugin's `package.json` (`engines: ">=18.18.0"` ), not an advisory floor, so an older runtime will refuse to install . On the account side, either a ChatGPT subscription — the free tier qualifies — or an OpenAI API key unlocks Codex . Auth is inherited, not re-entered. If you are already logged into Codex locally, the plugin reuses that state automatically — no second login . No Codex installed yet? `/codex:setup` can offer to run `npm install -g @openai/codex` when npm is present, or you can run that command manually first . The full official path is four slash commands run inside an active Claude Code session, not the two the viral clip advertises . Work through them in order: `/plugin marketplace add openai/codex-plugin-cc` `/plugin install codex@openai-codex` `codex` , the marketplace is `openai-codex` , which is why the reference reads `codex@openai-codex` .`/reload-plugins` `/codex:*` skills never surface, and the `codex:codex-rescue` subagent stays absent from `/agents` .`/codex:setup` `npm install -g @openai/codex` ; if Codex is present but not logged in, it prompts you to run `!codex login` from Claude Code .Verify before you rely on it: once `/codex:setup` completes without complaint, `/codex:review` should respond, and the `codex:codex-rescue` subagent should be visible in `/agents` . If either is missing, re-run `/reload-plugins` — a skipped reload is the most common reason the command surface fails to appear. The viral "two commands, 30 seconds" pitch covers only the marketplace add and the plugin install — steps 1 and 2. Skip `/reload-plugins` and the plugin is registered but dormant: its `/codex:*` slash commands never load into the running session, so nothing appears to work . Skip `/codex:setup` and the plugin never checks whether Codex is installed and authenticated — the failure is silent rather than a readable error, which is the worst kind to debug . Two behaviors deserve caution before you lean on this in real work: `Stop` hook to fire Codex on every Claude response. OpenAI and practitioners warn this can spin up a sustained Claude/Codex loop that drains usage limits fast — treat it as optional, not a default .Keep it current, too. `v1.0.6` , released July 8, 2026, removed shell expansion for git commands — a correctness and security fix for a plugin that shells out inside your repo. Earlier builds carried Windows-compatibility and app-server readiness regressions, so pin a recent version . After `/codex:setup` completes, the plugin exposes a focused set of `/codex:*` slash commands plus a `codex:codex-rescue` subagent visible in `/agents` . This is a command surface, not a general orchestration runtime — each command maps to one Codex action against your local checkout. | Command | What it does | Key flags | |---|---|---| `/codex:review` | Read-only Codex review of uncommitted changes or a branch diff | `--base` , `--wait` , `--background` | `/codex:adversarial-review` | Same target selection, but a skeptical pass that questions design decisions and assumptions, not just bugs | `--base` + optional focus text | `/codex:rescue` | Delegates investigation or a fix to the `codex:codex-rescue` subagent | `--model` (`spark` → `gpt-5.3-codex-spark` ), `--resume` , `--fresh` , `--effort` | `/codex:transfer` | Hands the active Claude Code session into a persistent Codex thread for longer delegation (added in v1.0.5) | — | `/codex:status` · `/codex:result` · `/codex:cancel` | Manage background jobs | — | One practical habit: before you lean on a blocking `--wait` , run background jobs and poll them with `/codex:status` and `/codex:result` , cancelling with `/codex:cancel` — `--wait` ties up your terminal until Codex returns . The concrete takeaway: install a pinned recent build, run `/codex:review` once to confirm the subagent responds, then wire the review commands into your ship checklist — you now have Claude planning and Codex reviewing in a single terminal . No. A ChatGPT free-tier account qualifies, and an OpenAI API key works as well . The plugin does not maintain its own credentials — it reuses whatever authentication your local Codex install already holds. If you are already logged into Codex on the machine, that auth is picked up automatically; if not, run `!codex login` from Claude Code once and you are set. No. Codex usage counts against your Codex limits, and Claude usage against Claude's — there is no shared quota, no cross-account auth, and no linked billing . The plugin is a convenience bridge that invokes your locally installed Codex CLI and app server; it does not prove the two agents share context, accounts, or permissions. Watch the optional "review gate" in particular, since a Claude/Codex loop can drain Codex limits fast. You most likely stopped at the two install commands and skipped `/reload-plugins` . Adding the marketplace and installing the plugin registers it, but the running session has not yet loaded the new skills. Run `/reload-plugins` , then `/codex:setup` , and the `/codex:*` slash commands plus the `codex:codex-rescue` subagent will show up in `/agents` . This is why the "two-command" framing is really four. `/codex:review` runs a standard, read-only Codex pass over uncommitted changes or a branch diff against a base ref such as `main` , supporting `--base` , `--wait` , and `--background` . `/codex:adversarial-review` runs a steerable, skeptical pass that actively questions design decisions and assumptions, using the same target selection plus optional focus text. Reach for the adversarial pass when you want a pre-merge second opinion, not a routine change summary. No. `openai/openai-developers-for-claude` is a separate plugin that bundles a Docs MCP server and API-setup skills, while `openai/codex-plugin-cc` exposes the Codex coding agent itself . They live in different repositories and serve different purposes; installing one does not give you the other. If your goal is code review and task delegation, `codex-plugin-cc` is the one you want.