claude-codex Skill Review: Running OpenAI's Codex CLI From Inside Claude Code

111 stars on SkillsMP and zero movement in the last seven days. That's not a bad sign — that's a skill that found its audience and settled. No hype spike, no churn. For a niche utility that bridges two competing AI coding tools, that kind of stable adoption is actually more reassuring than a trending burst would be.

Let me tell you what this skill does, whether it holds up under real use, and who should actually bother installing it.

What This Skill Does

The premise is simple: you're inside a Claude Code session and you want to invoke the local codex CLI — OpenAI's terminal agent — without switching contexts. This skill adds a /claude-codex slash command that shells out to a wrapper script (codex-run.sh) and runs Codex against your current workspace.

That's it. There's no credential handling, no API key juggling, no OAuth dance. It assumes Codex is already authenticated on your machine and just calls it. The skill is a thin orchestration layer, not an integration platform.

The author is sonichi, who maintains this under the Sutando project — an ambitious personal AI agent framework with voice, vision, and meeting participation features. The claude-codex skill is a small, focused piece of that larger ecosystem, but it stands alone fine.

Why This Actually Matters

Here's the real problem this solves: you're mid-session in Claude Code and you want a second opinion from a different model without breaking your flow.

Maybe Claude gave you an implementation and you want Codex to poke holes in it. Maybe you're debugging something subtle and you want to see if a different model spots what you're missing. Maybe you're on a team that standardized on Codex for certain tasks but you prefer Claude's interface for everything else.

The alternative is: open a new terminal, switch context, type out your Codex command, copy the output, come back to Claude. That's annoying enough that most people just don't do it. This skill removes that friction.

It also fills a gap that neither tool fills natively. Claude Code doesn't talk to Codex. Codex doesn't talk to Claude. This is a manual bridge, but it's a real one.

Key Capabilities Worth Knowing About

1. Uncommitted diff review

bash "$SKILL_DIR/scripts/codex-run.sh" --review --uncommitted -- "Prioritize bugs and missing tests"

This is the mode I'd use most. You've got staged or unstaged changes and you want a second model to look at them before you commit. The --review --uncommitted flag combination is a sensible default and the skill surfaces it prominently.

2. Base branch comparison

bash "$SKILL_DIR/scripts/codex-run.sh" --review --base main -- "Focus on regressions and security"

Useful for pre-PR review. Point it at your base branch and get a diff-scoped analysis. This is the kind of thing you'd normally set up a CI check for, but having it available interactively during development is genuinely useful.

3. Output capture

bash "$SKILL_DIR/scripts/codex-run.sh" --output-last-message results/codex-review.txt -- "Review the current workspace"

Being able to save Codex's last message to a file is underrated. You can pipe that into documentation, a PR description, or feed it back to Claude for synthesis. The fact that this is a first-class flag rather than a shell redirect afterthought suggests the author actually uses this workflow.

4. Sandbox defaults

The skill defaults to workspace-write sandbox mode and explicitly tells you not to use bypass flags unless the user asks. This is the right call. Codex can execute code, and running it with permissive settings in an automated context is how you get surprises. The guardrails here are conservative in a good way.

5. Graceful empty-argument handling

If you invoke /claude-codex with no arguments, it explains the available modes instead of failing silently or running something dumb. Small thing, but it means the skill is usable by someone who just heard about it and typed the command without reading docs.

Who Should Install This

Install it if: - You have both Claude Code and the Codex CLI set up and actively use both - You want to run second-opinion code reviews without context switching - You're building workflows that synthesize output from multiple AI models - You're in the Sutando ecosystem already and want the full skill set

Skip it if: - You don't have the Codex CLI installed and authenticated — this skill does nothing for you without it - You're looking for a way to get Codex access; this assumes you already have it - You want a deep integration where Claude and Codex actually collaborate in real time — this is fire-and-forget delegation, not a dialogue - You're on a locked-down machine where shelling out to external CLIs is restricted

The dependency on a locally installed and authenticated Codex CLI is the real filter here. If that's not already in your setup, this skill has zero value until it is.

How to Install

Add it to your Claude skills directory:

# Project-level
mkdir -p .claude/skills
git clone https://github.com/sonichi/sutando.git /tmp/sutando
cp -r /tmp/sutando/skills/claude-codex .claude/skills/

# Or globally
mkdir -p ~/.claude/skills
cp -r /tmp/sutando/skills/claude-codex ~/.claude/skills/

Then verify Codex is reachable:

codex login status

That's the only real prerequisite. The skill's own --check flag will tell you if something's misconfigured.

Concerns and Limitations

The wrapper script is a black box until you read it. The SKILL.md references codex-run.sh extensively but doesn't inline its contents. You need to go look at the actual script in the repo to understand what flags map to what Codex invocations. That's not a dealbreaker, but it means you should read it before trusting it in a production workflow. The GitHub source is public and readable — do that.

This is a thin wrapper, not a smart integration. Claude doesn't parse Codex's output, reason about it, or do anything intelligent with the response unless you explicitly feed it back. The skill delegates and optionally saves output. What you do with that output is entirely up to you. If you expected Claude to automatically act on Codex's suggestions, that's not what this is.

No streaming feedback in the Claude UI. Depending on how the script runs, you may be waiting for Codex to finish with no progress indicator visible in Claude's interface. For long Codex runs on large repos, that's a bad experience. Worth testing with your typical workload before committing to this in a time-sensitive context.

The Sutando project is alpha. The parent project is explicitly labeled early-stage. The claude-codex skill itself is simple enough that it's unlikely to break in weird ways, but if you're pulling updates from the repo, be aware that things can shift.

Codex CLI availability. As of mid-2025, the Codex CLI has had some access and availability turbulence. If OpenAI changes how the CLI authenticates or deprecates certain flags, this skill will need updates. The author has been responsive based on the star count, but it's a dependency you don't control.

Verdict

If you're already running both Claude Code and the Codex CLI, install this. It's low risk, low complexity, and it solves a real annoyance. The sandbox defaults are sensible, the empty-argument behavior is polite, and the output capture flag is genuinely useful for documentation workflows.

If you're not already using Codex, this skill gives you zero reason to start. It's a bridge, not an onramp.

The 111 stars at stable trend tells you this found its niche. It's not trying to be more than it is, and that's the right call for a utility skill. Worth the five minutes to set up if the prerequisites are already in place.

Find it on SkillsMP: https://skillsmp.com/skills/sonichi-sutando-skills-claude-codex-skill-md

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