OpenAI just built a tool for its biggest competitor. And it might be the smartest thing they’ve ever done.
By the end of this post, you’ll know how to set up two superpowered AIs reviewing your code inside your terminal. I’ll walk through the install and show you the three commands that changed how I build.
Why OpenAI Built a Plugin for Claude Code
I’m Charles Dove. I run CC Strategic Consulting and build in public over at Charlie Automates on YouTube. Let me tell you what’s actually going on right now.
The AI war isn’t about walling off anymore. It’s about embedding everywhere. OpenAI wants Codex to be used whether you’re inside their ecosystem or not. This signals a massive shift in how AI tools are being built right now. AI tools working together, not against one another.
If you can’t beat them, you join them.
This new plugin they dropped is an official OpenAI repo from GitHub. Not a community hack. Not an AI wrapper product. It already has 4,800+ stars on GitHub and it’s still blowing up.
The whole idea is simple. You can use Codex (OpenAI’s equivalent of Claude Code) inside your terminal in tandem with Claude Code as a separate command. You could have it review your code. You can run the rescue command. Or you could have it run the adversarial command, basically pitting these AI models against one another for a whole new perspective.
What this means for you: two super intelligent AI models working together to review your setup automatically. Claude can write your code. Codex can review it. One terminal, two brains.
People were doing this manually before. Having Codex open in one tab, Claude Code in another, ChatGPT on the desktop. Copy and paste all day. Now it’s all integrated.
Here’s the part nobody’s talking about. There’s a reason OpenAI made this free. Let’s be honest, it’s not charity. They want to be embedded in every developer’s workflow, even if it runs on Claude.
How to Install the Codex Plugin (Step by Step)
The install takes about 2 minutes. Here’s what you need to do.
Step 1: Add to the marketplace. Head to the Codex GitHub repo. Scroll down to the install section. Copy the marketplace command and plug it into Claude Code.
Step 2: Run the second command. It’s going to ask you about scopes. You can use it personally inside the workspace. Pick the first option.
Step 3: Reload the plugins. Run the reload command so Claude picks up the new plugin.
Step 4: Check your status. Run the status check command. This tells you whether it’s ready or not, what’s missing, and whether you need to log in.
More than likely, you haven’t logged into Codex yet. Open up a fresh terminal (without Claude running) and authenticate. If you have a $20/month ChatGPT subscription, that should be enough to test it.
Just log in, click continue, and you’re good to go. The “enable review gate” message should confirm it’s now enabled for your workspace.
That was the easy part. What it actually does will probably impress you.
Command #1: The Code Review
The first use case is the code review command. I ran it on a plugin I built called Go Viral Pro.
You run codex review and then pass the path from your plugin or project. Typically, you’d run the review command on a new feature you’ve built out. Something you kind of already have going.
I didn’t want to burn through my entire $20 subscription right off the bat. So I had it pick a third of the scripts and review just two features.
Here’s what happened. Codex read every line outside of Claude Code. It flagged security holes, edge cases (scenarios that might not normally happen), and any potential design flaws it found.
It’s kind of like you just picked up a senior dev overseas to review your code in 30 seconds. And it doesn’t cost you anything more than your pre-existing ChatGPT subscription.
The review gave me critical points, warnings, and informational notes. It looked at each script, rated severity, and gave me actionable items.
This is the polite version of using the plugin. The next command is where things get real.
Command #2: The Adversarial Review
The next command doesn’t just review. It actually tries to break your code. Kind of like the Aegis plugin we have set up inside the CC Strategic ecosystem for Claude Code. But this one uses a completely different AI model entirely, which comes with its own pros and cons.
The adversarial review command attacks your design decisions. It challenges your assumptions, finds failure modes, and helps you identify potential scale risks.
Let’s say you’re assuming your API calls or webhooks are going to arrive on time. Someone clicks a button on your site and you think it’s going to work immediately. What if it doesn’t? That’s the kind of stuff this command pinpoints. The kind of things that might break at 2 AM with your biggest client on board.
When I ran the adversarial review on my Go Viral Pro plugin, it found real issues. The scrapers for Instagram and YouTube had some problems with construction. Some arbitrary writes. Some of the logic could silently produce corrupted inputs. I’d actually seen this happen before, so the findings were legit.
It flagged data corruption risks where competitor data could bleed into other datasets. Not good. But now I knew about it before shipping to production.
The adversarial review gave me a verdict: “needs attention.” Along with specific critical findings, recommendations, and vulnerable impact areas for each file.
If you’re using Codex and want to get more extensive with it, the $20 subscription probably won’t be enough for heavy use. But for targeted reviews on specific features, it’s perfect.
Command #3: The Rescue Command (My Favorite)
This is the one that changed everything for me. What if you could hand off a bug to another AI and just keep working while it investigates?
That’s exactly what the Codex rescue command does.
If you’re stuck on a bug, don’t stop working. Just run the rescue command. You describe the problem, but it also takes context from the adversarial review above.
Here’s the workflow that makes this so powerful. Codex runs in the background while you’re building other things. Let’s say you’re on Phase 5 of a build and you ran the rescue command. You can keep building Phase 6, 7, or 8 while it does everything in the background.
You can run the codex status command anytime you want an update on where the rescue is at. And you can grab the full results with the codex result command.
This is how you find the exact root cause, the exact file and lines, and a concrete fix that you can hand over to your orchestration tool (like PAUL) to build on and fix in another phase.
Two agents working in parallel on one codebase, all integrated inside your terminal, running with Claude Code. That’s not just a demo. That’s how I actually work now.
The Full Workflow: Review, Attack, Rescue, Remediate
Let me walk you through the complete workflow I used.
1. Run the standard review. Get a baseline understanding of issues in your codebase. This catches the obvious stuff.
2. Run the adversarial review. This goes deeper. It attacks your assumptions and finds things that might break under pressure.
3. Run the rescue command on specific bugs. Let it investigate in the background while you keep building.
4. Feed everything back to your orchestration layer. I took all the context from Codex and fed it back to PAUL (my development orchestration tool). It created Phase 9 as a remediation plan with five sections to clean everything up.
5. Execute the fix. When you’re ready, just run the command and it starts executing.
I delegated bugs to one AI and kept chatting and building with another. We got root causes and all issues compiled together in literally less than 10 minutes.
What This Costs You
The plugin itself is free. The install takes 2 minutes. And the code review replaces what would cost you maybe $150 to $200 an hour with an external contractor.
If you have a ChatGPT subscription, there’s really no reason you shouldn’t set this up today. Even the $20/month plan works for basic reviews and targeted adversarial testing.
For heavier usage, you might want to upgrade your ChatGPT plan. But for most developers and builders, the standard subscription covers plenty of review cycles.
Why Two AI Models Beat One
Think about it this way. When Claude writes your code, it has blind spots. Every AI model does. Having a second model (from a completely different company with different training data) review that work catches things the original model never would.
It’s like having a second set of eyes, but those eyes think differently. They catch different patterns. They flag different risks.
The adversarial review is especially powerful because it’s actively trying to break what Claude built. That adversarial tension produces better code than either model working alone.
And the rescue command means you never have to stop building to deal with a bug. Background delegation. You keep shipping while Codex investigates.
FAQ
Do I need a ChatGPT subscription to use the Codex plugin? Yes. You need at least the $20/month ChatGPT Plus subscription. That gives you enough API access to run reviews and the rescue command. For heavy usage, you might want a higher tier plan.
Does the Codex plugin work with any version of Claude Code? It works as a Claude Code plugin, so you need Claude Code installed and running. The plugin installs through the marketplace and runs alongside your existing setup.
Will Codex modify my code? No. The review and adversarial commands only analyze and report. They don’t change your files. The rescue command investigates bugs but gives you findings and recommendations. You decide what to fix and when.
Can I run Codex in the background while Claude Code is building? Yes. That’s the whole point of the rescue command. Codex processes run in the background. You can keep building with Claude Code and check status whenever you’re ready. This is the feature that changed my workflow the most.
Is this safe? Am I sending my code to OpenAI? The Codex plugin runs through your ChatGPT subscription, which means your code is processed by OpenAI’s systems. If you’re working on sensitive client code, review OpenAI’s data handling policies first. For personal projects and side builds, it’s a non-issue.
What’s the difference between the review and adversarial commands? The review command is the “polite” version. It reads your code, flags security holes, edge cases, and design flaws. The adversarial command actively tries to break your code. It challenges your assumptions, finds failure modes, and identifies scale risks. Use review for general health checks. Use adversarial before shipping to production.
How long does a full review take? The standard review finished in about 30 seconds for a few files. The adversarial review took a bit longer. The rescue command runs in the background so timing depends on the complexity of your bug. Overall, all three tests completed in under 10 minutes for me.
Can I use this with other orchestration tools besides PAUL? Absolutely. The Codex output is just text. You can copy findings into any tool, any workflow, any project management system. PAUL happens to be what I use through Charlie Automates and CC Strategic, but the concept works anywhere.
Get Started Today
If you want to go deeper on Claude Code plugins, templates, and personalized guides, join CC Strategic AI on Skool. It’s a free community with 2,000+ members. No pressure to upgrade.
Inside the Skool community, I have a PDF playbook you can follow along with for this exact setup. It’s inside the Claude Resource Classroom.
If you want hands-on help setting up your Claude Code environment the same way I run mine, work with me 1-on-1. We’ll get your terminal, plugins, and workflows dialed in.
And if you’re an agency owner or business doing $2-10M+ looking for AI automation, book a call with CC Strategic. We build the systems that let you ship faster with fewer people.
I drop one to three long-form videos every week on Claude Code plugins, workflows, and guides over at @charlieautomates on YouTube. I only teach what actually moves the needle forward for your business.
These tools are absolutely nutty. Two agents working in parallel on one codebase. Background delegation. Root cause analysis in under 10 minutes. This is how building works now.
Take care, family.