<- all tokdocs

Claudex Loop: The Free Claude and Codex Review Skill, Explained

By , founder of Mavric and Apso

View on TikTok ->

Claudex is an open-source set of skills that makes Claude Code and OpenAI's Codex check each other's work: one model plans or builds, and the other provider reviews it. It is free and public on GitHub under the MIT license, so you do not need to DM anyone to get it. A TikTok promoting it gets the roles backwards for the most common setup and calls a four-month-old project "new."

The video, from @nateherkai, says "There's a new free skill that makes Claude and GPT-6 Astra work on the same task together" and asks viewers to DM "CLAUDEX" on Instagram for a copy. Here is what the project does.

What is Claudex Loop?

The project is chaseai-yt/claudex-loop on GitHub. The GitHub API shows it was created on June 5, 2026, four months before the video was posted on October 11, and had 3,133 stars and 340 forks. Its README says it "was previously grill-me-codex and crucible," and it keeps older skills in a legacy/ folder.

It ships four skills rather than one:

  • claudex-route for "A model recommendation or one scoped handoff"
  • claudex-loop for "Requirements, plan review, implementation, and final inspection"
  • codex-review
  • codex-build

The README states that "Claudex Route is independent of the Claudex Loop workflow."

The project credits Chase AI for Claudex Loop, Matt Pocock for the legacy interview skills, and Peter Steinberger for the Codex-as-builder pattern. The video names none of them.

Who plans, who builds and who reviews?

The video says "First, Claude writes the plan, then GPT-6 Astra audits it," and later "Astra builds the project, and Claude audits the implementation." The README's routing table depends on where you start:

  • Start in Claude Code: Claude writes requirements and the plan, Codex reviews the plan, Claude builds by default, and a fresh Codex session does the final inspection.
  • Start in Codex: Codex writes the plan, Claude reviews it, Codex builds, and a fresh Claude session inspects.

So in Claude Code, Claude builds by default, which reverses the video's claim. You can change it: "Choose either builder with builder=claude or builder=codex. The inspector follows the builder choice and always uses the other provider."

The principle the video describes is real. The repository's own description says "Whoever built it never grades it." The README adds a condition the video skips: "If the coordinator takes over fixes, those new edits need another independent inspection."

Does Claudex use GPT-6 Astra?

Only if you choose it. The README says: "Model choices remain configurable. Use Claude Fable 5.1 and GPT-6 Astra when selected and available on your accounts, or retain each CLI's configured model." The settings include reviewer_model, builder_model and inspector_model, each defaulting to "each CLI's configuration." The claudex-route section suggests "Luna may suit a focused fixture-generation task, Sol a bounded implementation, and Astra or Fable a difficult review."

Both named models exist. OpenAI announced GPT-6 Astra on September 3, 2026, with the API id gpt-6-astra. Anthropic's models overview lists Claude Fable 5.1 as claude-fable-5-1, with a 1M token context window and pricing of $10 per input MTok and $50 per output MTok. A maintained fork, ujconsulting/claudex-loop, defaults its review roles to gpt-6.1-sol/medium.

Does the review loop run until the plan is right?

No. The video says the models "keep going until the plan actually holds up." The loop has caps:

  • rounds / MAX_ROUNDS defaults to 5 plan-review rounds
  • MAX_FIX_ROUNDS defaults to 2 build-fix attempts
  • MAX_INSPECTION_ROUNDS defaults to 2 (an inspection plus one reinspection)

Each review ends in APPROVED, REVISE or BLOCKED, and the README says "exhausted round budgets are surfaced rather than converted to approval."

The authors are careful about what a pass means: "A clean structured result does not prove the model is right... Zero findings is valid; a large number of findings is not a quality score." They describe their one published run, 55 findings over five rounds on a CRM plan, as "an illustrative run, not a controlled benchmark of model pairings."

How to install Claudex Loop

The skills use the open Agent Skills format, a folder containing a SKILL.md file. Claude Code reads personal skills from ~/.claude/skills/ (docs), and Codex reads them from $HOME/.agents/skills (docs). The README copies the skills into both:

mkdir -p ~/.agents/skills ~/.claude/skills
cp -R skills/. ~/.agents/skills/
cp -R skills/. ~/.claude/skills/

It also offers a plugin install in Claude Code: /plugin marketplace add chaseai-yt/claudex-loop, then /plugin install claudex-loop@claudex-loop.

The video says you can "drop it straight into Claude or Codex." The README lists more setup: "Both CLIs must be installed and authenticated for the full cross-provider workflow. Python 3.10+ runs the shared adapter; no runtime pip packages or separate API keys are required." It suggests checking codex --version, codex login status, claude --version and claude auth status. Without both CLIs signed in, the cross-model review cannot run.

Does the video show Claudex running?

It does not. The plan and task screens in the video are motion graphics. The two real screen captures show other work: a ChatGPT session analyzing nutrition-brand advertising, and a Claude Code session running npm run build. Neither shows a reviewer, a verdict or a second model. The GitHub page in the opening shot is headed "Karpathy-Inspired Claude Code Guide," which belongs to a different project family.

Bottom line

Claudex Loop is a real, free, MIT-licensed project you can clone from GitHub today. When you start in Claude Code, Claude plans and builds while Codex reviews and inspects, which is the reverse of the video's build step. The model pairing is configurable, and the review loop stops after a fixed number of rounds.

Sources

Published October 11, 2026. Based on a TikTok by @nateherkai. Full video notes (markdown)