Claude Code Channels (MCP) now talks bidirectionally to Codex App Server in the same live session (first bridge)

[raysonmeng]

What variant of Codex are you using?

CLI

What feature would you like to see?

Just 24 hours after Anthropic dropped Claude Code Channels (research preview), I built the very first bridge that plugs MCP notifications directly into OpenAI’s fully open-source Codex App Server (JSON-RPC).

Key points:
• True bidirectional communication inside one single live session
• Full mid-execution injection works both ways
• Pure local MCP server — no cloud hop
• Most meta: the entire AgentBridge was co-written by Claude Code + Codex while they were using the bridge to build itself

Repo + live demo (Claude literally greets Codex and they co-author the code together):
https://github.com/raysonmeng/agent-bridge

I also just opened a symmetric discussion on the Anthropic side:
anthropics/claude-code#36871

This feels like the foundation for a universal Agent-to-Agent collaboration protocol.

Would love feedback from the Codex team or anyone experimenting with interop!

Additional information

• Repo + live demo (Claude and Codex greeting each other): https://github.com/raysonmeng/agent-bridge
• Screenshot of the demo (Claude & Codex collaborating in same session) attached below
• This bridge uses pure local MCP server + JSON-RPC, no cloud hop
• This could be the foundation for a universal "Agent-to-Agent MCP" protocol across providers

[ofekron]

I maintain Better Agent, which supervises native Claude, Codex, and Gemini coding sessions in one local workspace. This bridge is promising; the main protocol detail I would tighten is the phrase “one single live session.” For recovery and audit purposes, I would model it as one logical collaboration with two provider-native sessions, not one merged session.

Three invariants have mattered in our multi-provider event path:

1. Keep identities separate. Carry a bridge/collaboration ID plus Claude session ID and Codex thread/session ID. A reconnect or provider-side fork should not silently change the logical collaboration, and the shared ID should not erase native lineage.
2. Make injection replay-safe. Every cross-agent message needs a stable message ID, sender, per-sender sequence or cursor, causation ID, and explicit acknowledgement. After either app server restarts, resume from the last acknowledged cursor; an exact replay must be a no-op. Mid-execution delivery makes duplicate side effects much more dangerous than duplicate chat text.
3. Separate communication from authority. Receiving an MCP notification should establish what the other agent reported, not automatically authorize tool execution, file writes, or approval. The receiving session should apply its own permission policy and record whether the message was observed, accepted as context, or acted upon.

A useful restart test would kill each side independently after send-before-ack and after ack-before-checkpoint, then assert exactly-once projection at the UI/audit layer even if transport delivery is at-least-once. That is where bridges that look solid in a live demo usually reveal ambiguity.

It is source-available and free for non-commercial use; commercial use requires separate permission.

AI-assistance disclosure: this comment was drafted by Codex under the maintainer’s authorization and reviewed in Better Agent.