~/guides/LOOP.md
Claudex Loop: Claude Code plans, Codex reviews it round by round, code starts once it passes (free plugin, how to use)
Claude Code and Codex review each other's plan until APPROVED, then the other model inspects the code. Two install commands, how to start, limits.
What happens when a model from another company picks holes in the plan Claude Code wrote? Claudex Loop is a free, open-source plugin: Claude Code writes the plan, Codex reviews it round after round, Claude revises and sends it back until Codex says APPROVED, and only then does coding start; when the code is written, the model that did not write it inspects it. This is the full version of my LOOP reel: how the roles split, how to install it in Claude Code (two commands), how to start a loop, how to review an existing plan only, what it needs and when to skip it. Everything follows the official README and SKILL.md, checked on 10 November 2026.
01The result first: the plan went through five review rounds, problems fell from 26 to 0

Claudex Loop covers the part before code is written and the part after. You describe what you want in Claude Code and Claude writes the plan; then Codex, a model from another company, reviews it. Every round Codex returns a verdict: REVISE, APPROVED or BLOCKED. Claude revises and sends it back to the same Codex session until it passes or the round cap is reached. Code only starts once the plan passes, and when it is written, the model that did not write it inspects it.

- What it does: who writes, who reviews, who inspects.
- How to install it in Claude Code: two commands you can copy.
- How to start a loop, how to review an existing plan only, how to change the round cap.
- What it needs, when not to use it, and what I have not run myself.
02What it is: one writes, the other reviews, and the writer never grades its own work

Claudex Loop is a set of skills for Claude Code and Codex, MIT-licensed and free. The repo description ends with "Whoever built it never grades it."

- Start in Claude Code: Claude asks the questions and writes the plan, Codex reviews it, Claude builds by default, and a fresh Codex session inspects at the end.
- Start in Codex: the reverse, Codex plans and Claude reviews.
- To have Codex build, add builder=codex; the inspector is always the other one.
- The plan lives in PLAN.md; every round's findings, decisions and results go into PLAN-REVIEW-LOG.md.
03Step 1: install it in Claude Code (two commands)

Type these two lines in Claude Code:
/plugin marketplace add chaseai-yt/claudex-loop /plugin install claudex-loop@claudex-loop
- After installing, the commands are /claudex-loop:claudex-loop (the full workflow), /claudex-loop:codex-review (Codex reviews a plan only) and /claudex-loop:codex-build (Codex builds).
- Codex can use it too: clone the repo and copy the whole skills/ folder into ~/.agents/skills/ (Codex) and ~/.claude/skills/ (Claude Code); the README has the macOS / Linux and Windows PowerShell commands.
- Open a new session so the skills show up.
04Step 2: start a loop

Open Claude Code in your project and ask for the full workflow, followed by the feature you want:
/claudex-loop:claudex-loop [the feature you want, e.g. add CSV import to this project, plan it and implement it]
It reads your code first, lists its assumptions for you to confirm, asks only about decisions that change the outcome, then writes PLAN.md. Then Codex reviews:
- Codex reads the plan and the relevant code, backs every finding with evidence, and ends with REVISE, APPROVED or BLOCKED.
- REVISE: Claude weighs each finding, changes the plan where it is warranted, writes down why when it is not, and sends it back to the same Codex session.
- APPROVED: the plan's path and content hash are recorded; change the plan later and it needs another review.
- The default cap is 5 rounds; if it still has not passed, it shows you the open findings instead of pretending it passed.
If you already have a plan and only want it reviewed:
claudex this plan, mode=review, plan=[PLAN.md or docs/your-plan.md], rounds=3
05Step 3: code only after the plan passes, then the other one inspects

- The builder works from PLAN.md and records the commit before building; a build delegated to Codex needs a clean checkout (README).
- It runs the agreed proof commands, then opens a fresh session where the other model inspects every changed file.
- Findings get fixed and re-inspected; by default up to 2 fix rounds and 2 inspections (the first one plus one after fixes).
- At the end you get the diff, the proof results, what the inspection covered and anything still unresolved.
/claudex-loop:claudex-loop [the feature you want], builder=codex
06Good to know

- Claude Code and Codex CLI both have to be installed and logged in, plus Python 3.10+. No separate API key, but it uses both accounts' limits; if the reviewing side runs out, it does not quietly switch models, it waits for the quota to reset.
- Skip it for small changes. The official SKILL.md says "not for trivial edits".
- APPROVED is a model's verdict, not a guarantee. The README says a clean result does not prove the model is right, and a large number of findings is not a quality score. Still run your tests.
- The reviewer is read-only: Codex runs in a read-only sandbox with plugins and MCP servers switched off; Claude can only read and search files.
- The repo is an open-source project by Chase AI; the README footer links his paid community. This guide has no partnership with it.
- I have not run the full loop on my own project: the 26 → 0 image is the official example, and the commands here come from the README and SKILL.md.
Sources
What should the next reel cover?
I test Claude and AI tools every day and turn the useful ones into reels. Follow @vivianlin.tech to catch each new guide.