claude-codex Skill Review: Running OpenAI Codex From Inside Claude Code Actually Works

111 stars on SkillsMP and zero movement in the last week. That's not a red flag — that's a skill that found its audience and stopped needing to market itself. The claude-codex skill from sonichi (part of the Sutando project) has been sitting quietly at that number, which tells me it got real installs from people who actually needed it, not hype installs from people chasing the trending tab.

Let me tell you what it actually does, whether it's worth your time, and where it falls short.

What This Skill Does

In plain terms: it gives Claude Code a /claude-codex slash command that shells out to your locally installed codex CLI. You type a prompt, the skill wraps it in the right codex invocation, runs it against your current workspace, and optionally saves the output to a file.

That's it. There's no credential handling, no API key juggling, no OAuth dance. It explicitly assumes Codex is already authenticated on your machine. If you have codex login working, this skill just routes prompts through it.

The implementation lives in a shell script (codex-run.sh) that the SKILL.md calls with different flags depending on what you're doing — review mode, exec mode, output capture. It's a thin wrapper, not a reimplementation.

Why This Actually Matters

Here's the problem it solves: you're paying for both Claude and Codex (or you have access to both through work), and right now they live in completely separate contexts. If you want a second opinion from Codex while you're mid-session in Claude Code, you have to stop, open a terminal, remember the right flags, run the command, copy the output back. That friction is small but it adds up.

This skill collapses that into /claude-codex review this function for edge cases and you're done.

The more interesting use case — and the one I think justifies the install — is second-opinion analysis on diffs. When you're about to merge something and you want a model that wasn't involved in writing the code to look at it fresh, being able to fire codex review --uncommitted from inside your current Claude session without context-switching is genuinely useful. Different models have different blind spots. Using both isn't paranoia, it's redundancy.

The Sutando project that this skill comes from is building toward a broader "personal AI agent" vision — voice, screen sharing, multi-machine coordination, the whole thing. This skill is one small, self-contained piece of that. You don't need to buy into the Sutando vision to use the skill. It stands alone.

Key Capabilities Worth Knowing

1. Uncommitted diff review The --review --uncommitted flag is the one I'd use most. It points Codex at your staged and unstaged changes and asks it to look for bugs and missing tests. This is the highest-signal use case for the skill — you're getting a fresh set of eyes on exactly what's about to ship.

2. Base branch comparison --review --base main lets you scope the review to everything that diverged from main. Useful in longer-running feature branches where you want to audit the whole thing before opening a PR, not just the last few commits.

3. Output capture The --output-last-message flag saves Codex's last response to a file. This matters more than it sounds. If you're running Codex in a CI-adjacent workflow, or you want to log review results somewhere persistent, having that as a first-class flag rather than > output.txt tacked on the end is cleaner.

4. Sandbox defaults The guardrails section specifies workspace-write sandbox or stricter by default, and explicitly says not to use bypass flags unless the user asks. That's a sensible default. You don't want a delegated Codex run making unexpected filesystem changes outside your repo.

5. Workspace anchoring It keeps Codex pointed at $PWD with -C "$PWD" unless you tell it otherwise. Sounds obvious, but getting this wrong would be annoying to debug. The fact that it's explicit in the guardrails means someone thought about it.

Who Should Install This

Install it if: - You actively use both Claude Code and Codex CLI and find yourself switching between them - You want a quick second opinion on diffs without leaving your Claude session - You're building automation around code review and want a way to pipe Codex results into files programmatically - You're already in the Sutando ecosystem

Skip it if: - You don't have Codex CLI installed and authenticated. This skill does zero setup for you — it's a bridge, not an installer - You're looking for a deep integration where Claude and Codex share context, pass artifacts back and forth, or have a conversation. That's not what this is - You only use one of the two tools. The value is entirely in having both - You're on a team where Codex isn't part of the standard toolchain. Installing a skill that shells out to a binary your teammates don't have is a support headache waiting to happen

How to Install

Add it to your Claude skills directory:

# Project-level (recommended)
mkdir -p .claude/skills
cd .claude/skills
git clone --filter=blob:none --sparse https://github.com/sonichi/sutando.git
cd sutando
git sparse-checkout set skills/claude-codex

Or just copy the skills/claude-codex directory into ~/.claude/skills/ for global availability. Make sure codex is in your PATH and codex login status returns something sane before you try the skill.

Concerns and Limitations

The dependency is opaque. The skill assumes codex CLI exists, is authenticated, and behaves a certain way. If OpenAI changes the CLI's flag interface — which has happened before — this skill breaks silently or noisily depending on how the shell script handles errors. There's no version pinning here.

The shell script is the actual implementation. I'd want to read codex-run.sh before trusting this in any sensitive environment. The SKILL.md shows the interface but the script is where the real behavior lives. It's open source, so go read it — but don't skip that step.

No output streaming back into Claude's context. From what I can tell, Codex runs as a subprocess and the output either goes to stdout or a file. Claude doesn't get to reason about the Codex output in the same turn unless you do something extra to pipe it back. That limits the "second opinion" use case somewhat — you get the review, but Claude can't immediately synthesize it with its own analysis unless you paste it back in.

Sutando project scope. The parent project is ambitious — voice control, meeting bots, multi-machine agent coordination. That's exciting but it also means this skill is a small piece of a larger system that's still in alpha. The skill itself is self-contained, but if you're evaluating the long-term maintenance trajectory, you're betting on a project that's doing a lot of things at once.

111 stars is solid but not massive. This isn't a skill with a large community stress-testing edge cases. If you hit a weird interaction, you're probably filing the first issue about it.

Verdict

If you're using both Claude Code and Codex CLI regularly, install it. The use case is real, the implementation is simple enough that there's not much to go wrong, and the guardrails are sensible. It won't change how you work — it'll just remove a small but consistent friction point.

If you're only using one of the two tools, or you're evaluating this hoping it does something more sophisticated than "run codex CLI with these flags," move on. It's not trying to be more than it is, and that's fine.

The honest summary: this is a well-scoped shell wrapper with a clean interface. It does one thing. That thing is useful if you need it.

SkillsMP page: https://skillsmp.com/skills/sonichi-sutando-skills-claude-codex-skill-md

GitHub: https://github.com/sonichi/sutando/tree/main/skills/claude-codex