agent-bridge

A tiny local MCP relay that lets two (or more) MCP-speaking agent CLIs — e.g. Claude Code and Grok Build — talk to each other through shared channels, instead of a human copy-pasting between two terminals.

Both CLIs already speak MCP, so the "bridge" is just a shared MCP server they both connect to. This is that server. It's generalized: any future MCP CLI (Cursor, Codex, etc.) can join the same channel with zero changes here.

• Transport: stdio (each CLI spawns its own copy of the server).
• Storage: local append-only JSONL, one file per channel, in ./data/. Nothing leaves your disk. Survives restarts.
• Deps: @modelcontextprotocol/sdk, zod. Node 18+.

Tools it exposes

Tool	Purpose
bridge_post(channel, from, body)	Append a message to a channel
bridge_read(channel, since?, limit?)	Read messages with seq > since
bridge_wait(channel, since, timeout_ms?)	Block until a newer message lands — event-driven, wakes the instant the other agent posts
bridge_channels()	List channels + last seq

Each message gets a monotonic seq per channel. Remember the highest seq you saw and pass it as since next time so you only get new messages.

Watch mode — get "triggered" on every message

An MCP server is request/response: it can't push a new turn into an idle CLI session, so there's no way to wake a session that's just sitting there. The way to make a session react the instant a message arrives is to keep it parked in bridge_wait, which is now event-driven (fs.watch) and returns within milliseconds of a post — no busy-polling.

There are two ways to do this.

A) In an interactive session (no extra process)

Tell your agent to loop:

Watch the zero-ui-debate channel. Call bridge_wait(channel="zero-ui-debate", since=<lastSeq>). When it returns with messages, handle them and bridge_post your reply, then immediately call bridge_wait again with since set to the highest seq you just saw. If it returns timedOut, just call it again. Keep looping until I say stop.

The session now wakes and acts on every new message on its own — the human only starts the loop once. (Each bridge_wait blocks up to timeout_ms, default 60s, max 10min; the loop simply re-arms it, so the watch is continuous.)

B) As a background daemon — watch.js

watch.js encodes the trigger so you don't paste a loop prompt: it fs.watches a channel and runs a command of your choice for every new message. Point that command at a headless agent and you get a hands-off auto-responder.

# Whenever someone other than "claude" posts to zero-ui-debate,
# spawn a headless Claude to read context and post a reply:
node watch.js zero-ui-debate --from claude -- ./examples/respond-with-claude.sh

• --from <me> skips your own posts, so the responder never triggers on its own replies (no infinite loop). Always set it.
• Each fire gets the message via env (BRIDGE_CHANNEL, BRIDGE_SEQ, BRIDGE_FROM, BRIDGE_BODY, BRIDGE_TS) and as JSON on stdin.
• Messages are handled one at a time (next waits for the current command to exit), and progress is persisted next to the channel, so a restart resumes instead of replaying. Use --replay to process existing history first.

examples/respond-with-claude.sh is a ready template that calls claude -p with this server attached so it can bridge_post the reply itself — adapt the prompt, model, or swap in grok for the other side.

Note the boundary: a daemon spawning a fresh responder per message is the closest you can get to "trigger the session." An MCP server still cannot inject a turn into an already-running, idle interactive session — nothing can, for a generic CLI. Option A keeps one session live; option B starts a worker per message.

Setup (one time)

cd /Users/prakshaljain/OS-1/agent-bridge
npm install
node smoke-test.mjs   # optional: verifies post/read round-trip

Then add the server to both CLIs. The entry is in .mcp.json.example:

{
  "mcpServers": {
    "agent-bridge": {
      "command": "node",
      "args": ["/Users/prakshaljain/OS-1/agent-bridge/server.js"]
    }
  }
}

• Claude Code: claude mcp add agent-bridge -- node /Users/prakshaljain/OS-1/agent-bridge/server.js (or paste the block into a project .mcp.json / ~/.claude.json), then restart the session.
• Grok Build: it reads .mcp.json / claude_desktop_config.json automatically — add the same block and restart. (grok.com/connectors works for the web app too.)

After both restart, each CLI will have bridge_post / bridge_read / bridge_wait / bridge_channels.

How the debate loop runs

1. Claude posts its position to channel zero-ui-debate (from: "claude"). (Already seeded — see debate/.)
2. In the Grok terminal, paste debate/GROK-BRIEF.md. Grok reads the channel, reads debate/STRATEGY-POSITION.md, and posts its rebuttal (from: "grok").
3. Claude calls bridge_wait(channel="zero-ui-debate", since=<lastSeq>), reads Grok's turn, and replies. Repeat until converged.
4. Both write article drafts into articles/.

No human relay needed once both are connected — each side parks in bridge_wait (see Watch mode) and is triggered the moment the other posts.

Folders

• server.js — the MCP relay
• watch.js — background daemon that runs a command on every new message (watch mode B)
• examples/ — ready-to-edit responder templates for watch.js
• data/ — channel logs (gitignored)
• debate/ — the briefing prompt + the seeded strategy position
• articles/ — shared drafting workspace for the campaign deliverables