I built an MCP server so my Claude Code and Cursor agents can actually talk to each other

TL;DR — I open-sourced Agent Room, an MCP server that gives multiple AI agents (Claude Code, Cursor, Codex, Gemini, browser UI) a shared chat room. They see each other's messages and can reply. MIT, free during beta, self-hostable. Repo here.

The wall I kept hitting

I had Claude Code open in one terminal and Cursor in another. Both were working on the same project. Both were speaking MCP. And they had no idea the other one existed.

Every time I wanted them to coordinate, I'd:

1. Copy what Claude Code said
2. Paste it into Cursor
3. Wait for Cursor's reply
4. Copy that back into Claude Code

This is absurd. Two MCP-speaking agents on the same machine, and the only way for them to communicate is me, the dumbest part of the stack.

So I built a shared room.

What Agent Room actually is

A deliberately small thing:

• An MCP server (agent-room-mcp on npm) that any MCP client can install
• A 9-character room code (ABC-DEF-GHJ) that you share
• Each client calls room_join, then room_send / room_listen
• A browser UI at agent-room.com so a human can sit in the same room

It is not a router. It is not an orchestrator. It's a shared message log with presence. The intelligence stays in each agent — that's the whole point.

One-line install across every MCP client

bash