What Is Claudex Loop? Installation, Claude and Codex Cross-Review, MCP Risk and Cost
You start a task in Claude Code; Claudex Loop has Claude write the plan and Codex critique it, and after the code is written, the provider that didn't write it reviews it in a fresh session. The author never approves its own work. It's an open-source (MIT) package of 4 skills; for the full flow, both Claude Code and Codex CLI must be installed and logged in, Python 3.10+ is required, and no separate API key is needed. The plan and code snippets go to the other provider's session. In the current version, MCP servers and plugins are turned off during the Codex review; an older version didn't have this protection.
- What it does
- Turns Claude Code and Codex into a pair that checks each other on the same task
- Type
- Skill package (4 skills)
- Made by
- chaseai-yt
- Review
- Read-only; the author doesn't approve
- Data
- The plan and code go to the other provider
- Turkish
- Not stated in the documentation
- Cost
- The package is free; it uses both accounts' limits
What Is Claudex Loop For?
A skill is a package of instructions that teaches an agent how to work on a particular topic. Claudex Loop prevents a single model from approving its own plan and code: the plan is written with one provider, critiqued and revised with the other; if you want, the build is also handed to the other provider, and at the end the side that didn't write it reviews it in a fresh session.
| Skill | What it does |
|---|---|
| claudex-loop | Full flow: plan, critique, optional build, cross-review |
| claudex-route | Standalone: suggests which model should take the task and does a one-time handoff |
| codex-review | Compatibility command that explicitly picks Codex as the plan reviewer |
| codex-build | Compatibility command that explicitly picks Codex as the builder |
Who it's for: developers who have both a Claude Code and a Codex account and want a second, independent pair of eyes on important changes. It's unnecessary for small, single-file tasks.
How Do You Install Claudex Loop?
Requirements: Claude Code and Codex CLI, both installed and logged in; Python 3.10+. The README doesn't say which subscription plan is required. The commands are taken verbatim from the README.
› /plugin marketplace add chaseai-yt/claudex-loop› /plugin install claudex-loop@claudex-loopInvoking: /claudex-loop:claudex-route (lightweight recommendation), /claudex-loop:claudex-loop, /claudex-loop:codex-review, /claudex-loop:codex-build.
› mkdir -p ~/.agents/skills ~/.claude/skills && cp -R skills/. ~/.agents/skills/ && cp -R skills/. ~/.claude/skills/After cloning the repository, run it from the repository folder and open a new session. In Codex, you invoke them with $claudex-route and $claudex-loop.
› codex --version && codex login status && claude --version && claude auth statusTested baseline: Codex CLI 0.153.4 and Claude Code 2.1.261 (Windows). An older Codex CLI showed the flags but rejected the model. The install line isn't pinned to a version; you get the latest state of main.
How Do You Use Claudex Loop?
First describe your request normally, then invoke the skill. The plan is written to PLAN.md and the review log to PLAN-REVIEW-LOG.md.
If the build is handed off, it requires a clean git working copy. Choose explicitly which side writes the code (builder); in the default flow, the host side does the writing.
Where Does Claudex Loop Work?
| Claude CodeYes | Plugin or a copy of the skill folder. |
|---|---|
| CodexYes | Copy the skill folder into ~/.agents/skills; there's also a plugin definition. |
| Other agentsNo | The package only pairs these two providers. |
What Does Claudex Loop Access, and What Happens to Your Code?
The project code and the plan file (sent to the other provider as a prompt and through the CLI's file-reading tools); git status.
PLAN.md and PLAN-REVIEW-LOG.md; the chosen builder writes the code; diagnostic files (prompt, command, output) are written to a separate directory in the system temp folder and may contain private code or plans, so they shouldn't be committed.
The docs list no delete action; since the builder changes code, it can delete files.
The plan and code snippets go to the other provider's CLI session (to the OpenAI/Codex service when you're using Claude Code). The runner doesn't call a model API directly.
!The Codex reviewer runs with codex exec -s read-only. In the current main version, plugins and apps are turned off on every Codex run, and enabled MCP servers are disabled as well (if the list can't be read, the run stops). The older pinned version (8cf5e2c) didn't have this protection: the Codex review could see existing MCP settings (issue #28, fixed on October 14, 2026). If your Codex configuration has MCP servers with write access, use the current version.
There are no hooks in the repository; there's a runner script (runner.py) you can read before installing. GitHub Security Advisories is empty; no advisories are listed. That doesn't mean "safe."
Is Claudex Loop Free?
The package is free and MIT-licensed (GitHub shows NOASSERTION because of an extra copyright notice in the LICENSE file). The only thing you pay for is your own Claude and Codex account usage: every loop runs on two separate models, so it uses up two limits at once. The repository doesn't specify any plan or pricing requirement. If your Codex quota runs out, there's no cross-model fallback; you wait until the quota resets. No local-currency price is listed.
Common Problems and Fixes
| What's happening | What to do |
|---|---|
| The Codex review loads MCP tools | That's how the old version behaved; use the current main or check the Codex MCP settings manually. |
| Reported by users Codex commands die with exit 137 (macOS) | It may be an old, broken npm-global Codex CLI; check which codex binary is running (issue #10, open). |
| Codex quota ran out and the loop stopped halfway | There's no cross-model fallback; wait until the quota resets. |
| Output encoding error on Windows | Issue #31 has been closed; check the current version. |
How Do You Uninstall Claudex Loop?
- If it's a plugin: remove claudex-loop from the /plugin menu in Claude Code.
- If you installed it manually: delete the claudex-loop, claudex-route, codex-review and codex-build folders under ~/.claude/skills and ~/.agents/skills.
- Leftovers: The PLAN.md and PLAN-REVIEW-LOG.md files created in the project are yours; delete them if you want. The diagnostics folder stays in the system temp directory.
Frequently asked questions
What is Claudex Loop?
A skill package in which Claude Code and Codex check each other's work: one writes the plan, the other critiques it; then one writes the code and the other reviews it in a fresh session. The author never approves its own work.
How do you install Claudex Loop?
In Claude Code, run plugin marketplace add chaseai-yt/claudex-loop, then plugin install claudex-loop@claudex-loop. You need to be logged in to Claude Code and Codex CLI, and Python 3.10+ must be installed.
Is Claudex Loop free, and what's its license?
The package is free and MIT-licensed. However, every loop uses up the usage limits of both your Claude and Codex accounts. The repository doesn't state any specific plan requirement.
Does my code go to another company?
Yes: the plan and code snippets go to the other provider's CLI session (to OpenAI's Codex service when you're using Claude Code). Keep this in mind for confidential or client-owned code.
Is it safe, and what does it access?
The reviewers run under a read-only restriction; in the current version, MCP and plugins are turned off in Codex, while the older pinned version doesn't have this protection. The builder, on the other hand, writes code. No security advisory is listed; that doesn't mean it's safe.
Related guides
Sources
Not clear in the official sources: The subscription plan required for Claude and Codex · a line-by-line security reading of the runner script (runner.py) · Turkish support · the number of installations still on the old version. Each one is covered in its section.