Claudex Loop: Make Claude Code and Codex Check Each Other's Work

When one model writes the plan and then checks its own work, it tends to agree with itself. Claudex Loop splits those jobs between two different AI providers, Claude and Codex. Its own tagline: whoever built it never grades it. Everything below comes from the repo's README, checked on October 3, 2026.

Here's the flow when you start in Claude Code. Start in Codex and the roles flip.

1. 1Recon. Claude reads your existing code and docs (or researches the assumptions on a new project) and shows you an assumptions list with sources.
2. 2Requirements. It asks you the questions that change the outcome, then writes PLAN.md with acceptance checks and the commands that prove it works.
3. 3Plan review. Codex reads the plan and the relevant code and sends back findings backed by evidence. Claude decides what to change, updates the plan and sends it back. This repeats until Codex says APPROVED or BLOCKED, or the round cap is reached (5 rounds by default).
4. 4Build and inspect. Once you authorize the build, the builder (Claude by default when you start in Claude Code) writes the code, the proof checks run, and a fresh Codex session inspects the final changes.

Every finding, decision and model used goes into PLAN-REVIEW-LOG.md, so you can see what was caught and why the plan changed. A request to review a plan never counts as permission to build.

First check that both tools are installed and signed in. These are the checks the README lists:

Then, inside a Claude Code session, add the plugin:

Run the full workflow with /claudex-loop:claudex-loop. The plugin also adds /claudex-loop:claudex-route, /claudex-loop:codex-review and /claudex-loop:codex-build. Using Codex as your home base instead? The README covers copying the skills folders into ~/.agents/skills and calling $claudex-loop.

• builder: claude or codex. The inspector is always the other provider.
• rounds: how many plan-review rounds before it stops. 5 by default.
• mode=review: start from a plan you already wrote instead of from scratch.
• inspect: on by default. Turning it off is allowed, and it gets logged.
• Fix and inspection caps: 2 build-fix attempts, and an initial inspection plus one re-inspection, by default.

The same repo includes Claudex Route, a lighter skill that recommends which model should handle a task and can do one handoff when you ask. It runs on its own and doesn't need Python. Call it with /claudex-loop:claudex-route, for example: "Recommend a second opinion on this plan before we build."

• Use it on work where a wrong plan is expensive: features that touch many files, data migrations, auth or payments.
• Skip it for one-line fixes. Several review rounds across two models use up usage on both accounts.
• Reviews are kept on a short leash: Codex reviews run in a read-only sandbox, and Claude reviews only get file reading and search tools. Builds handed to the other model need a clean git checkout.

Now you just ship code that's already been double checked. Inside the Claude Code Club we share the review and testing workflows we actually run, and help each other set them up. It's nine dollars a month at https://www.skool.com/claudecodeclub/about. Everything on this page works without it.