OpenAI Embeds Codex into Claude Code: Inside the New Plugin

Someone posted a screenshot showing Claude Code terminals with commands prefixed by /codex: and labeled (codex), captioned “Introducing Codex Plugin for Claude Code”.

The official OpenAI repository openai/codex-plugin-cc was created on March 30 and amassed 2,300 stars in less than a day.

The initial commit includes the line Co-authored-by: Codex <[email protected]>, indicating that OpenAI used Codex itself to write the plugin.

Heavy AI‑tool users are excited because the plugin creates the first officially sanctioned collaboration channel between two competing AI assistants.

Previously, moving code from Claude Code to Codex required copying, switching windows, pasting context, and retrieving results, which introduced significant friction; the plugin eliminates this overhead.

The plugin provides an /codex:adversarial-review command that differs from the regular /codex:review by not only finding bugs but also questioning design decisions and suggesting safer or simpler alternatives. The README lists suitable scenarios such as authentication logic changes, data migrations, and infrastructure scripts that are costly to fix after deployment.

/codex:adversarial-review --base main challenge whether this was the right caching and retry design

Users can specify the focus of the adversarial review, e.g., challenging a caching design, which is akin to asking the AI to “find where I might be wrong” rather than merely “check for errors”.

No other mainstream AI coding tools currently offer a comparable standardized command.

Using /codex:rescue, tasks can be delegated to Codex with optional --background execution. Progress can be checked with /codex:status and results retrieved via /codex:result:

/codex:rescue --background investigate the flaky integration test
/codex:status
/codex:result

This enables seamless task flow between the two AIs; /codex:result returns a Codex session ID that can be resumed in the native Codex CLI with codex resume <session-id>, preserving continuity across tool switches.

The /codex:rescue command also accepts a model flag. For example, spark is a shorthand for gpt-5.3-codex-spark, a lightweight, fast model suitable for quick runs without deep reasoning, allowing users to balance cost and performance.

Technically, the plugin does not embed any Codex runtime; it simply calls the locally installed Codex CLI, reusing the user's existing ~/.codex/config.toml or project‑level .codex/config.toml configuration, resulting in essentially zero‑configuration for existing Codex users.

Claude Code users who have not used Codex need only a ChatGPT subscription (the free tier suffices) or an OpenAI API key and Node.js 18.18+. Running /codex:setup checks the environment and provides setup guidance.

Claude Code’s plugin marketplace underpins this integration. The plugin can be installed with:

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

The /plugin marketplace add command takes a GitHub repository path, meaning any GitHub repo can be turned into a Claude Code plugin, highlighting the marketplace’s openness.

OpenAI’s decision to publish the plugin through this mechanism, rather than waiting for Anthropic, demonstrates the marketplace’s accessibility and suggests that the ecosystem will organically fill gaps.

An experimental “Review Gate” feature can be enabled with /codex:setup --enable-review-gate. It triggers an automatic Codex review before Claude Code replies, blocking Claude’s output if issues are found. However, the documentation warns that this can cause an infinite loop of back‑and‑forth between the two AIs, rapidly consuming tokens, and recommends enabling it only in sessions that are actively monitored.

Project address: https://github.com/openai/codex-plugin-cc