Claude Code plugin for Codex

Use Claude Code from Codex to review code or delegate tasks.

The mirror image of openai/codex-plugin-cc, which does the same thing in the other direction. If you work in both, run both.

Delegation goes through your local claude CLI, so it reuses the auth, config, MCP servers, and CLAUDE.md you already have. There is no separate account, no API key to paste, and no second runtime.

What You Get

Mode	Ask for it like	What it does
task	"delegate this failing test to claude"	One claude run. Read-only unless you ask for edits.
review	"have claude review my working tree"	Correctness, security, and edge cases over local git state.
adversarial review	"have claude poke holes in this approach"	Challenges the design and its assumptions, not just the implementation.

Codex picks the mode from your wording. After any review it stops and asks which findings you want fixed — it never auto-applies them.

Requirements

• Codex CLI v0.117 or later (for the plugin system)
• Claude Code CLI: npm install -g @anthropic-ai/claude-code
• A Claude subscription or an Anthropic API key, logged in once via claude

Install

codex plugin marketplace add https://github.com/linkesch/claude-plugin-codex
codex plugin add claude@linkesch-claude

From a local clone instead:

git clone https://github.com/linkesch/claude-plugin-codex
codex plugin marketplace add ./claude-plugin-codex
codex plugin add claude@linkesch-claude

Verify it registered:

codex plugin list | grep claude@linkesch-claude

Usage

Just ask, in words. There are no slash commands — Codex routes to the skill on intent.

Delegate a task

delegate this failing test to claude

Runs read-only: Claude investigates and reports back, without touching files. Ask for the edit explicitly to let it write:

have claude fix the failing test

Review

have claude review my working tree

Reviews local git state for correctness bugs, security issues, and edge cases. Untracked files aren't in git diff, so the skill checks git status first — an empty diff with new files present is not "nothing to review".

Adversarial review

have claude poke holes in this approach is this the right design? why shouldn't this ship?

Uses the prompt in prompts/adversarial-review.md. Default stance is skepticism: it hunts for the strongest reasons the change should not ship, prioritising expensive and hard-to-detect failures — trust boundaries, data loss, rollback safety, races, version skew. It reports only, never edits, and an empty finding list is a valid answer.

Follow-ups

keep going dig deeper into the second finding

Continues the last Claude conversation in that directory via claude -p -c.

Typical Flows

Review before shipping

have claude review my working tree

then, once you've picked what matters:

have claude fix findings 1 and 3

Second opinion on a design

have claude poke holes in this approach, focus on the retry logic

Hand over a problem you're stuck on

delegate this to claude: the worker deadlocks under load, figure out why

Troubleshooting: "Not logged in · Please run /login"

In the Codex app and normal interactive use, this does not come up — the plugin works out of the box against your existing Claude login.

You can hit it when Codex runs claude inside its sandbox, most commonly under codex exec, which defaults to a read-only sandbox with no approvals. On macOS, Claude Code keeps its OAuth token in the keychain, and seatbelt blocks keychain access, so the call fails even though you are perfectly logged in.

Run it outside the sandbox. An API key does not help — the sandbox blocks outbound network too, so the call fails either way.

This is the one asymmetry with codex-plugin-cc, which has no such problem: Codex keeps its auth in a plain ~/.codex/auth.json that Claude Code can read from inside its own sandbox.

FAQ

Do I need a separate Claude account for this plugin?

No. It shells out to the claude CLI you already logged into.

Will it use my existing Claude config?

Yes — same config, same MCP servers, same CLAUDE.md, same permissions. The plugin adds no configuration of its own.

Which model does it use?

Whatever your claude CLI defaults to. Name one to override:

ask claude with opus to review this

Can it edit my files?

Only when you ask. Default is read-only; edits require explicit phrasing, which maps to --permission-mode acceptEdits.

Why are there no slash commands?

Codex plugins expose skills, not slash commands. Ask in plain language.

Why is there no background mode or job tracking?

codex-plugin-cc needs a companion daemon because it drives Codex over JSON-RPC. This direction doesn't: claude -p prints to stdout and claude -p -c resumes the last conversation in a directory, so the session, job, and transfer machinery collapse into flags. Runs are foreground. Large reviews take a while.

Development

The plugin is four content files: a marketplace manifest, a plugin manifest, a skill, and the prompt/reference the skill pulls in. No build, no dependencies.

Changes to source do not go live until you reinstall — Codex serves the plugin from a cache under ~/.codex/plugins/, even on an unchanged version:

codex plugin add claude@linkesch-claude

Bump version in plugin.json and add a CHANGELOG.md entry for anything user-visible.

License

MIT — see LICENSE.

Not affiliated with, endorsed by, or sponsored by OpenAI or Anthropic.

"Claude", "Claude Code", "Codex", and their logos are trademarks of their respective owners. The Claude glyph in plugins/claude/assets/claude.svg is Anthropic's, used unmodified and only to identify which tool this plugin delegates to. The MIT license above covers this project's own code and docs, not that mark.