// AI-to-AI message broker

Claude sessions
that talk to each
other

Five agents. One bridge.
Questions go one way, answers come back — in real time, while the humans sleep.

view on github
Five Claude agents exchanging messages over claude-bridge Animated diagram showing five Claude Code and Claude Desktop sessions — frontend, backend, research, database, and tests — wired together on a single message bridge. Question packets travel along the wires and reply packets return, illustrating real-time AI-to-AI communication over MCP. frontend claude-code backend claude-code research desktop db claude-code tests claude-code

    claude-bridge is a local MCP server for multi-agent Claude Code and Claude Desktop workflows. Each Claude session registers on the bridge, then sessions can ask each other questions, reply with context, and share a scratchpad — turning isolated AI agents into a coordinated team that exchanges information over the Model Context Protocol.

    1. 01

      register

      Each Claude session announces itself on a local broker. One line, lifecycle hook.

    2. 02

      ask

      One agent calls ask(to, question). Its turn blocks until an answer comes back.

    3. 03

      reply

      The target agent sees a notification, replies with context. Both keep working.

    4. 04

      scratchpad

      Shared notes any agent can read or append. No more re-asking what's already known.

    Zero deps. ~600 lines of Node. Localhost only.
    MCP-over-SSE for the CLI, MCP-over-stdio for Claude Desktop.