intercom-mcp
acunningham-ship-it/intercom-mcp
A message bus / coordination layer for AI agent fleets — MCP server letting Claude Code (or any MCP client) sessions ask, answer, and broadcast. No daemon, no port; shared SQLite + a live fleet dashboard.
Instalação
npx convez-ai acunningham-ship-it/intercom-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
intercom-mcp
A coordination layer for AI agent fleets. An MCP server that lets agent sessions (Claude Code, Codex, or any MCP client) on one machine talk to each other: ask a question and block for the answer, fire-and-forget a message, broadcast status, and watch the whole fleet move in real time. No daemon, no port.
How it works
No daemon, no port. Each agent session spawns its own stdio instance of
server.js; every instance reads/writes one shared SQLite database
(~/.local/share/intercom/intercom.db, WAL mode), so all sessions see the same
bus. Presence is pid-based: an agent is "online" while its server process is
alive, and dead rows are pruned automatically.
Claude session A ── server.js ─┐
Claude session B ── server.js ─┼─→ intercom.db (SQLite WAL)
Claude session C ── server.js ─┘
Tools
| Tool | What it does |
|---|---|
join | Register under a name (e.g. project name) + see who else is on. Call first. |
who | List online sessions: name, role, cwd, last activity. |
send | Fire-and-forget message to one agent, or broadcast (omit to). |
ask | Send a question and block until answered (default 60s, max 240s). Times out gracefully — the question stays queued. |
reply | Answer a question / respond to a message by #id. Unblocks a waiting ask within ~1s. |
inbox | Fetch unread messages. wait_seconds long-polls until something arrives. |
history | Re-read recent traffic (optionally filtered to one agent). |
Delivery & notifications
Delivery is durable: anything you miss is waiting in your inbox. For real-time
notification without polluting your session's transcript, arm a watcher:
- Monitor (recommended): run
monitor-watcher.js --me <name>under your agent's background-monitor mechanism. It emits one line per new message, so fresh mail shows up as a clean event instead of being typed into your context. - Status line:
claude-code-statusline.jsputs an unread count in the status bar, fully out of the way. - Pull / blocking:
inboxlong-polls (wait_seconds), andaskalready blocks until the other sidereplys (~1s), so request/response needs no watcher at all.
Broadcasts are pull-only by design: they don't nudge every session, so a busy fleet doesn't spam everyone. Directed messages are the ones that notify.
Watch the fleet
node watch.js is a live, read-only dashboard of the bus: who's online, the recent
message flow, and which questions are still open. It's how you actually see a fleet of
agents coordinate. Add --follow to stream it. Example output:
┌─ Intercom Fleet Status
├─ Online Agents (4)
│ boss · worker-2 · worker-3 · worker-4
├─ Open Questions (1)
│ #57 worker-2 → boss: "which lane owns the config file?"
├─ Recent Message Flow
│ #58 boss → worker-2 "take the parser lane, file-disjoint from w3"
│ #59 worker-3 → boss "lane done, 12/12 green"
└─ (read-only; ^C to exit)
Ops
- Register it with any MCP client by pointing the client at
node /path/to/intercom-mcp/server.js(stdio). Use an absolutenodepath so it doesn't depend onPATH. Every session that registers shares the same bus. - Test:
npm test(spawns 3 real server processes on a temp DB; no API key needed). - Override the DB path:
INTERCOM_DB=/path/to.db. - Retention: messages + read-receipts older than
INTERCOM_RETENTION_DAYS(default 7) are pruned onjoin; set0to disable. - Needs node >= 22.5 (for
node:sqlite). - Waits are capped at 240s to stay under MCP tool timeouts; raise
MCP_TOOL_TIMEOUTif you push the caps.