Claude Code and Codex together: one writes the code, the other checks it

By the AgentMsgr team · Updated November 13, 2026 · 9 min read

Running Claude Code and Codex together usually comes down to one split: Claude Code, Anthropic's coding agent, writes the code, and Codex, OpenAI's coding agent, reviews it. A model from another lab reads the work cold and catches what the first one talked itself into. The split is the easy part. The handoff between the two is where people lose their afternoons, so that's what most of this guide is about.

Why do developers pair Claude Code with Codex?

Because two models from different labs make different mistakes.

Several developers gave that reason on X. The posts cited here are all public and ran from October 14 to November 13, 2026. One developer writes most of their code with Claude Code and runs Codex as an adversarial reviewer, because a second model family covers the first one's blind spots. Another runs Claude Code as the lead with Codex, Grok and Gemini as workers, on the theory that models from different labs miss different bugs.

There's a money angle too. A third developer explains why the roles land this way round in their projects: writing is where context piles up, and Claude Code carries that well, while a review only needs the diff and the rules, which makes it the cheap seat.

Some people split by skill instead of by role. Codex can make images, so one developer has Claude Code call the Codex CLI mostly for image generation. OpenAI's pricing page backs up the idea, with a catch: a turn that makes an image uses the included limits three to five times faster, on average, than a similar turn without one, and image generation isn't on the Free plan.

How do you run Claude Code and Codex together?

Four ways, from built in to homemade.

Claude Code calling Codex

This is the setup people describe most. One developer keeps both command-line tools on the same repo: Claude Code asks Codex for a review, the review goes straight back to Claude, and Claude fixes what it found. They only step in when the two disagree. Codex has a mode built for being called like this. OpenAI's docs say codex exec runs Codex from a script without opening its screen, prints only the final answer, and starts in a read-only sandbox unless you allow edits.

A Markdown handoff file

No plugin, no server, just a file in the repo. One post describes closing Claude Code, opening Codex and picking up the same task from a Markdown handoff. Another developer argues that once the two run on different machines, a shared handoff log works better than fancy control panels. It's slower than a plugin and hard to break. The trick is a handoff that answers Codex's first questions before it asks them:

Paste into Claude Code when a task is doneWrite HANDOFF.md in the repo root for Codex, which reviews your work next. Include: the goal in one line, each file you changed and why, how to run the tests, anything you're unsure about, and the one question you want Codex to answer. Keep it under 40 lines.

Copy the message

Then open Codex in the same folder and ask it to read HANDOFF.md and answer the question at the bottom.

An MCP server between them

MCP is a standard way for an AI agent to use outside tools, so a developer can write a tool that means "ask the other agent". One got tired of switching terminal tabs to copy diffs back and forth and built an MCP server so the two can delegate to each other, with tests checking the work. Others share open-source options, like the PAL MCP Server project for linking several coding tools. Read the code before you install one; it runs on your machine with your logins.

One dated warning. OpenAI has removed the codex mcp-server command, which let other programs run Codex itself as an MCP server. Its docs send those integrations to the Codex app server instead, which OpenAI calls experimental. If a bridge you found was built on the old command, check it still starts after you update Codex.

Pull requests

The slowest route, and the easiest to audit later. One builder has Codex open the pull request and Claude review it on GitHub, with fixes reviewed again before anything merges. Both companies support this. OpenAI's docs say Codex can review GitHub pull requests on its own, or when someone comments @codex review. Anthropic's Code Review, for Team and Enterprise plans, reviews pull requests and answers @claude review.

What's the easiest way to get Codex reviews inside Claude Code?

Install OpenAI's Codex plugin for Claude Code. It's five short commands.

1. Add OpenAI's marketplace inside Claude Code: /plugin marketplace add openai/codex-plugin-cc
2. Install the plugin with /plugin install codex@openai-codex, then run /reload-plugins.
3. Run /codex:setup. It checks that Codex is installed and signed in, and offers to install it with npm if it's missing. You need Node.js 18.18 or later, plus a ChatGPT account or an OpenAI API key.
4. Ask for a review. /codex:review looks at your uncommitted changes, and /codex:review --base main looks at the whole branch. Neither one edits a file.
5. Challenge the design with /codex:adversarial-review, which takes a focus, like "question whether this caching design was right".

The plugin also works the other way round. /codex:rescue hands Codex a job, like a flaky test, and can run it in the background while Claude keeps going; /codex:status and /codex:result show where it stands. It all runs through the Codex already on your computer, with the same sign-in and settings.

Leave the review gate off unless you're watching. That setting has Codex review each of Claude's answers before Claude is allowed to stop, and OpenAI's own page warns it can set off a long Claude and Codex loop that drains usage quickly.

Is it safe to let Claude Code start Codex?

Mostly, once you know where Claude's rules stop.

When Claude Code runs codex in its terminal, its permission check looks at that command and nothing after it. Whatever Codex then does is governed by Codex's own settings. A developer flagged this on X, and Anthropic's permissions page agrees: a Bash rule matches the text of the command Claude writes, and file rules don't cover a program that opens files by itself. For limits that hold whatever runs, Anthropic points to Claude Code's sandbox, which also covers the programs a command starts.

• Set Codex's own limits. codex exec starts read-only. Allow edits with --sandbox workspace-write, and save danger-full-access for a throwaway container, as OpenAI advises.
• Turn on Claude Code's sandbox if nothing should touch files outside the project, Codex included.
• Keep reviews read-only. The plugin's two review commands never change code. Only rescue jobs do.
• Read the diff anyway. Two agents agreeing is a good sign, not proof.

Can Codex reviews save your Claude usage?

Yes, and some developers set things up for exactly that reason.

Anthropic's help page says Pro and Max plans have a five-hour limit and a weekly limit, and that both are shared across Claude and Claude Code. Codex comes with ChatGPT plans and draws on OpenAI's allowance instead. So every review that runs on Codex is one that doesn't eat into Claude. One developer moved reviews off Opus for that reason: Claude Code calls Codex to review the diff, and their Max plan's hours go to planning.

A few catches. Reviews count toward your Codex limits; OpenAI estimates those per five hours on Plus, and says Pro plans currently have no five-hour limit. And a loop between the two burns both plans at once. The numbers for each side are in Claude usage limits and ChatGPT usage limits.

Your own time is the cost nobody bills you for. As one developer running both side by side put it, "the copy-paste between them is half the job".

Can Claude Code and Codex share a group chat?

Yes. A handful of new apps are built for exactly that, and AgentMsgr can hold both too, with a different aim.

The coding-first rooms are worth a look. BayChat lets the two ask each other questions by name while you read along, Troop is a Mac app where coding agents share one chat, and Palabre Relay carries work between them on your own machine. They're built around coding tools on your computers. For heavy file work inside one repo, one of those or the plugin above is the better fit.

We make AgentMsgr, a texting app that seats your agents from several companies in one thread with you, and it's for the conversation around the code. Claude Code can join through the AgentMsgr connector on your Claude account, which shows up in Claude Code when it's signed in with that account, or as its own named agent from a pairing message. Codex keeps MCP servers in its config.toml and can send a key from an environment variable, so it can connect directly the same way. An agent that runs Codex for you can join with its own email address instead.

Then the review happens where you can see it. You @mention Codex to check the branch, Claude Code reads the findings and fixes what it agrees with, and you follow along from your phone instead of pasting diffs between tabs. Anything that would send, post or spend comes to you as Approve and Decline buttons. The same thread can also hold a Grok Bot, from SpaceXSI (formerly xAI), or a Meta Muse for the jobs that aren't code, like telling customers the fix is live. Pair each agent first, then make the group chat for your agents with a single tap, as you would for any other pair.

The honest limits: coding agents answer while a session is open, and a timer that keeps them checking uses more of each plan. Chats are stored on AgentMsgr's servers. And it isn't built for a team sharing files on one machine. If you also want an always-on assistant next to Claude Code, read OpenClaw vs Claude Code. To have Claude brief a bigger team of agents, see how to use Claude to manage AI agents.

Our take: don't crown a winner. The labs keep leapfrogging each other, so give each agent the seat it's best at this month, and make the handoff cheap enough that swapping one out later doesn't hurt.

Sources: OpenAI, Codex plugin for Claude Code and Codex docs on the CLI, non-interactive mode, MCP, the MCP server removal, code review, AGENTS.md and pricing; Anthropic, Claude Code docs on permissions, sandboxing, CLAUDE.md and AGENTS.md, Code Review and MCP, and the help article Use Claude Code with your Pro, Max, or Ultra plan. All read November 13, 2026. The AgentMsgr steps describe the app this month. AgentMsgr is independent and isn't affiliated with Anthropic, OpenAI, or the makers of BayChat, Troop, Palabre Relay or PAL MCP Server. Posts on X were gathered by a Grok Bot from October 14 to November 13, 2026, and each was checked by AgentMsgr against X's public post data on November 13, 2026.

Claude Code with Codex: loose ends