Codex Claude Bridge

Codex Bridge

Claude Code and Codex CLI talk to each other through one markdown file in your project.
No server, no hooks, no plugin. You read along.

Quick start

1. Add talkto.md to your project root:

2. Open claude and codex in that project, or in any of its git worktrees.

3. Tell one of them what to discuss:

4. Tell the other one check talkto.md. From then on they keep going on their own.

How it works

The rules sit at the top of talkto.md, so each agent learns the protocol by reading the file. There is nothing to install.

1. An agent appends a message that ends with Waiting on:  — .
2. It runs the watch command from the file in the background.
   - Claude Code runs it with Monitor and is woken when it prints.
   - Codex CLI runs it in a background terminal and keeps polling, so its turn stays open.
3. The watch exits when a complete message addressed to that agent appears. The agent reads it, replies, and waits again.
4. Waiting on: nobody — done ends the conversation.

The watch compares against the agent's own last message, not a line count. A reply that lands before the watch starts is still caught, and a half-written one waits until its Waiting on: line exists. Every git worktree resolves to the main checkout's talkto.md.

Message format

Optional: skip naming the file

Add one line to AGENTS.md, which Codex reads, and to CLAUDE.md, which Claude reads:

Then discuss this with codex is enough.

Limits

- An agent waits only while its turn is running. If Esc, an error or a usage limit cuts a turn short, say check talkto.md.
- Both agents must run on the same machine.
- Anything appended to talkto.md reaches both agents as a message.
- Agents read only the latest messages, but a long file still costs context. Start a fresh file when a topic is…