OpenAI Just Built a Plugin for Their Competitor's Tool. I Installed It.

OpenAI built a plugin for Anthropic's tool.

Read that sentence again. OpenAI — the company that makes ChatGPT, GPT-5.4, and Codex — shipped an official, Apache 2.0 licensed plugin that runs inside Claude Code. Their direct competitor's coding terminal.

It's called codex-plugin-cc. It's on GitHub under the openai org. It's real.

I installed it today.

🤯 Why Would They Do This?

The cynical answer: Claude Code is eating their lunch. By early 2026, Claude Code hit roughly $2.5 billion in annualized revenue and accounts for about 4% of all public GitHub commits — around 135,000 per day. That's a lot of developers who aren't using Codex.

The strategic answer: if you can't beat the IDE, join it. Developers were already doing this manually — writing code with Claude, then copy-pasting diffs into Codex for a second opinion. OpenAI saw that pattern and thought, "what if we just... made that a slash command?"

The honest answer: it's good business. Every /codex:review someone runs inside Claude Code is a Codex API call that counts against their OpenAI usage. OpenAI gets paid. The developer gets a better review. Claude Code becomes stickier because it now has superpowers from two AI companies. Everyone wins, sort of.

It's like Pepsi putting a vending machine inside a Coca-Cola factory. Except the factory workers actually want both drinks, and they've been sneaking Pepsi in through the back door anyway.

📦 What the Plugin Actually Does

Six slash commands:

The first two are the ones that matter. The adversarial review is particularly interesting — it's not just "find bugs." It asks things like "what happens to this under 10x load?" and "what assumption are you making that could be wrong?" It's the reviewer who isn't trying to be polite.

/codex:rescue is for when Claude gets stuck on something and you want a completely different brain to take over. Different model, different reasoning, different blind spots.

🛠️ How I Installed It

Step 1: Install Codex CLI

npm install -g @openai/codex

Took 4 seconds. Requires Node.js 18.18+.

Step 2: Check Authentication

codex login status

Mine said Logged in using ChatGPT. If yours doesn't, run codex login first. You need a ChatGPT subscription (even free tier works) or an OpenAI API key.

Step 3: Add the Marketplace

This is where it gets slightly messy. The README says to run /plugin marketplace add openai/codex-plugin-cc inside Claude Code. That works if you're using the CLI version. I'm on the VSCode extension, and the /plugin command wasn't recognized.

So I did it manually. Cloned the repo into Claude Code's marketplace directory:

cd ~/.claude/plugins/marketplaces
git clone https://github.com/openai/codex-plugin-cc.git openai-codex-plugin-cc

Then registered it in ~/.claude/plugins/known_marketplaces.json:

{
  "openai-codex": {
    "source": {
      "source": "github",
      "repo": "openai/codex-plugin-cc"
    },
    "installLocation": "path/to/openai-codex-plugin-cc",
    "lastUpdated": "2026-04-10T00:00:00.000Z"
  }
}

And enabled the plugin in ~/.claude/settings.json:

{
  "enabledPlugins": {
    "codex@openai-codex": true
  }
}

Restart Claude Code. Done.

If you're on the CLI version, the three-command approach from the README should just work:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins

🧠 Why Two AIs Are Better Than One

This is the part that actually matters.

Claude and Codex reason differently. Not "one is better, one is worse" — they literally see different things. Here's what the benchmarks show:

Claude writes longer, more documented, more architecturally thoughtful code. Codex writes tighter, faster, more production-oriented code. Claude catches design problems. Codex catches runtime problems.

Using both is like having an architect and a structural engineer review the same building plan. The architect says "this room layout doesn't flow well." The engineer says "this wall can't support that load." You need both opinions. Separately, each misses what the other sees.

The workflow that's trending among developers right now:

1. Claude builds — writes the code, handles the architecture
2. /codex:review — practical bug hunt, performance issues, security holes
3. /codex:adversarial-review — "what kills this in production?"

Three passes. Two brains. More bugs caught than either could find alone.

⚠️ What It Can't Do

A few things worth knowing before you get too excited:

It's not a Claude replacement when your plan gauge runs out. The plugin runs inside Claude Code. If Claude's rate limit is exhausted, the plugin doesn't work either. The plugin needs Claude to be alive to invoke it.

If your Claude gauge is dry, just use Codex directly in your terminal:

codex "fix the login bug in this project"
codex review

Codex CLI works independently. It uses your OpenAI quota, not your Claude quota. That's your escape hatch.

Codex doesn't know your project rules. It doesn't read your CLAUDE.md. It doesn't know your commit message conventions. It doesn't follow your team's linting preferences. Whatever Claude knows from your project configuration, Codex is blind to.

If you're delegating a task with /codex:rescue, include the rules in the prompt:

/codex:rescue Fix the auth bug. Korean commit messages, 
tests required, match existing code style.

It has its own rate limits. Every Codex call counts against your OpenAI usage. There's no /codex:usage command yet (someone filed an issue for it). If you're on the free ChatGPT tier, you'll hit limits faster than you'd like.

🔮 What This Means

A year ago, the idea of OpenAI building tooling for Anthropic's product would have been absurd. These companies are direct competitors. They're fighting for the same developers, the same enterprise contracts, the same benchmark rankings.

But developer workflows don't care about corporate rivalries. Developers use what works. And right now, what works is Claude for building and Codex for reviewing. OpenAI saw that pattern emerging organically and decided to own it rather than fight it.

This is the beginning of something. Not "AI companies becoming friends" — but AI tools becoming interoperable in ways their creators probably didn't plan for. Today it's a review plugin. Tomorrow it might be shared context protocols, unified billing, or cross-model debugging sessions.

Or maybe OpenAI just really wants those API calls. Either way, I'm keeping it installed.

2026.04.10