Backed by Y Combinator (F26)

Codex and Claude Talk: How to Make Codex and Claude Code Talk to Each Other

To make Codex and Claude talk to each other, connect your ChatGPT and Claude plans in Reason's Settings › Models, start a Session on a Claude or OpenAI model, and ask it to have a helper on the other model review or answer: the helper works in the same repository and branch, and its reply comes back into your conversation. It works in both directions in the Reason web app or Mac app, with no terminal and each model on the plan you already pay for, and if you'd rather stay in Claude Code on your computer, OpenAI's Codex plugin for Claude Code lets Claude Code hand reviews and tasks to Codex.

Reason runs in the Mac app and on the web.

  1. 1
    Settings › ModelsConnect your ChatGPT and Claude plans.
  2. 2
    In the message boxPick a Claude or an OpenAI model for the Session.
  3. 3
    Ask for a second opinionThe other model's reply comes back into the same Session.

A second opinion before you ship.

Make Codex and Claude Code talk in 4 steps

  1. Open Reason at reasonmachines.com, or in the Reason app for Mac from reasonmachines.com/download.
  2. In Settings › Models, connect your ChatGPT plan with Sign in with ChatGPT and your Claude plan with Connect Claude; API keys work too.
  3. Start a Session on your repository and pick a Claude or an OpenAI model in the message box.
  4. Ask for the other model by name, for example "have a helper on GPT-6 Luna from my ChatGPT account review this diff". It works on the same branch, and its reply comes back into your Session.

The detailed walkthrough with screenshots is below, followed by the other ways people connect the two agents.

Make Codex and Claude talk in Reason, step by step

  1. Open Reason on the web or in the Mac app

    Sign in at reasonmachines.com, or download the Reason app for macOS from reasonmachines.com/download. Both open the same Sessions. Sending work in Reason needs the 3-day trial, a paid plan or credits, also when it runs on your own subscriptions.

  2. Connect your ChatGPT and Claude plans

    Open Settings › Models. Connect a ChatGPT Plus, Pro or Business plan with Sign in with ChatGPT (or Codex sign-in), and a Claude Pro or Max plan with Connect Claude; OpenAI or Anthropic API keys work too. Connected accounts are listed by provider, as below.

    Reason Settings › Models with Claude and ChatGPT subscriptions connected, next to OpenAI and Google API keys
  3. Connect GitHub and start a Session

    In Customize › GitHub, connect the repositories Reason may work in. Then start a new Session, name the repository and describe the task. Reason works on its own cloud computer, or on your Mac through the Mac app.

  4. Pick the first model in the message box

    The model picker in the message box lists each connected account's models, grouped by provider. Pick a Claude model to have Claude write the change, or an OpenAI model to start with Codex's models.

    The model picker in a Reason Session's message box, listing Reason, ChatGPT, Claude, Google and OpenAI
  5. Ask the other model to review or answer

    Say it in plain words and name the model, for example "have a helper on GPT-6 Luna from my ChatGPT account review this diff and tell me what it would change". Reason starts a helper on that model in the same checkout and branch, and its reply comes back into your Session, where Claude can answer it or apply the fixes; the helper's row shows which model and account it ran on. Started on an OpenAI model? Ask for a Claude helper the same way.

    A Reason Session on Claude Opus 5.5 that starts a helper on OpenAI's GPT-6 Luna and quotes its reply in the same conversation
  6. Keep going when a plan runs out

    Each model runs on the account you picked and keeps that provider's limits. Hit your limit? Reason keeps going on your other connected accounts or Reason credits, and says so in the conversation.

What people tried to make Codex and Claude talk

Copy-paste relayCodex plugin for Claude CodeMCP bridgesReason
How it worksYou paste one agent's plan or diff into the other and carry the answer backOpenAI's Claude Code plugin adds slash commands that run Codex reviews and hand tasks to Codex from inside Claude CodeAn MCP server (for example agent-link-mcp) lets one agent start another agent's command line as a collaboratorOne Session on one model starts a helper on the other model; the reply comes back into the Session
DirectionWhichever way you carry itClaude Code asks CodexThe host agent asks the ones it startsBoth ways: Claude can ask an OpenAI model, and the other way round
SetupNoneInstall the plugin in Claude Code; needs a ChatGPT subscription (Free included) or an OpenAI API key, and Node.jsInstall and configure an MCP server on your computerConnect both plans in Settings › Models, in the web or Mac app
Where it runsWherever each agent runsYour computerYour computerReason's cloud computer, or your Mac
CostsEach agent's own planCounts toward your Codex usage limitsEach agent's own planEach model on the plan you connect; Reason needs the trial, a plan or credits
Good forA quick one-off second opinionCodex reviews without leaving Claude CodeAgents consulting each other on one machineBoth models working the same repository and branch, with no terminal

There's no built-in link between OpenAI's Codex and Anthropic's Claude Code: they are separate agents with separate context. These are the ways people connect them today, each fair for a different job.

Orchestrating Claude Code and Codex

Once Codex and Claude talk, give each a role: one plans, one writes, one reviews, and the work hands off between them on the same branch. In Reason the Session you're in is the orchestrator, and each helper it starts on the other model gets the same checkout, branch and permissions.

  1. Planner: start the Session on the model you trust to plan, for example Claude, and ask it to write the plan and the acceptance checks before any code. Then ask for a helper on an OpenAI model to challenge the plan, and settle the open questions before the change starts.
  2. Builder and reviewer: one model writes the change and runs the tests, the other reviews the diff. For example "have a helper on GPT-6 Luna from my ChatGPT account review this diff for bugs and missing tests"; the review comes back into the Session, where the first model answers it or applies the fixes before the pull request opens.
  3. Handoff: when a task fits the other model better, or you want a fresh pair of eyes on a stuck bug, ask the Session to hand that part to a helper on the other model with what's done and what's next. The helper works on the same branch and reports back, and you see which model and account each step ran on.
  4. At a limit: each model runs on the account you picked and keeps that provider's limits. When one account hits its limit, Reason keeps going on your other connected accounts for that model or on Reason credits, and says so in the conversation.

Keep the rules both models follow in the repository: Codex reads AGENTS.md and Claude Code reads CLAUDE.md, and Reason reads both. Ask for one reviewer per change rather than a long back-and-forth; two reviews that disagree are a signal to look at the code yourself.

The same roles work with the command-line agents on your own computer: Reason Network lets your live Claude Code and Codex sessions message each other, so one can plan or review for the other without you copying text between terminals.

How Reason differs

Copy-paste makes you the messenger, and the plugin and MCP bridges run on your computer and start from one agent. In Reason, both models work in the same checkout and branch, the reply lands in the conversation you're already in, and the work keeps running on Reason's cloud computer when your laptop is closed.

Already run Claude Code and Codex on your own computer and want those live sessions to message each other? Reason Network connects them, both ways and across computers, and it's free.

Connect other agents

Questions

Can Codex and Claude Code work together?

Yes. They are separate agents with no built-in link, but you can connect them: in Reason, a Session on one model starts a helper on the other in the same repository and branch, and the reply comes back into your conversation. On your own computer, OpenAI's Codex plugin for Claude Code lets Claude Code send reviews and tasks to Codex, and Reason Network lets live Claude Code and Codex sessions message each other.

Can Codex and Claude Code work on the same repo?

Yes. In Reason, the helper on the other model works in the same checkout and branch as your Session, so Claude and an OpenAI model review and edit the same code. Running both command-line agents yourself on one repository also works; give each its own branch or worktree so their edits don't collide.

Do Codex and Claude share context?

Not automatically: each model has its own context, and neither reads the other's conversation. In Reason, the Session hands the helper the task and what it needs to know, the helper reads the same files, and its reply comes back into the Session. Repository instructions help both: Codex reads AGENTS.md, Claude Code reads CLAUDE.md, and Reason reads both.

Is it free?

Partly. Connecting your ChatGPT and Claude plans adds no Reason fee for those models, which keep their own plans and limits, but sending work in Reason needs the 3-day trial, a paid plan (Pro is $50 a month) or credits. Reason Network, for live Claude Code and Codex sessions on your own computer, is free.

How is this different from the Codex plugin for Claude Code?

The plugin runs Codex from inside Claude Code on your computer, one way. Reason runs both models in one Session on your connected plans, in both directions, in the web or Mac app, on its own cloud computer. You can use both.

Do I need an API key?

No. Connect your ChatGPT plan with Sign in with ChatGPT and your Claude plan with Connect Claude in Settings › Models. API keys work too if you'd rather pay per use.

Is Reason made by OpenAI or Anthropic?

No. Codex is a product of OpenAI and Claude Code a product of Anthropic. Reason is built by Reason Machines and is not affiliated with or endorsed by either.

More guides