| name | intercom |
| description | Always active — coordinates with other Claude Code agents via intercom MCP. Automatically checks for messages, announces presence, and communicates to avoid conflicts. Triggers on ANY task start and periodically during work. |
Intercom — Inter-Agent Communication
Purpose
Coordinate with other Claude Code instances working in parallel. Avoid file conflicts, duplicate work, and wasted effort by communicating proactively.
MANDATORY — On Every Task Start
Before doing ANY work, run these two calls:
mcp__intercom__who() → discover who else is active on this projectmcp__intercom__peek() → check if anyone sent you a message
If other agents are active, consider sending a brief message about what you're about to work on.
When to Communicate (Proactively)
Send a message (mcp__intercom__send) when:
| Situation | Example message |
|---|
| Starting work on a file | "Je commence à modifier src/auth/middleware.ts" |
| About to create a PR or push | "Je vais push sur feature/auth, heads up" |
| About to create a git worktree | "Je crée un worktree pour feature/billing" |
| Finished a task that unblocks others | "Auth middleware terminé et mergé, vous pouvez pull" |
| Found a bug or issue affecting others | "Attention: la migration DB 042 est cassée, ne pas merger" |
| Need information from another agent | "Qui gère le rate limiter ? J'ai besoin de contexte" |
When to Check Messages
Call mcp__intercom__peek():
- At task start (mandatory)
- Before modifying shared files (package.json, schema, configs)
- Before git operations (push, merge, rebase)
- Periodically during long tasks (every ~10 tool calls)
- Before claiming work is done
How to Reply
When you receive a message:
- If it asks a question →
mcp__intercom__reply(message_id, "your answer")
- If it's informational →
mcp__intercom__ack(message_id) after reading
- If it warns about a conflict → adjust your work accordingly, then reply
Scope
who(scope="project") → agents on the same project (default)who(scope="all") → all agents everywheresend(to="all", ...) → broadcasts to same-project agentssend(to="x7k2", ...) → direct message to a specific agent (works cross-project)
Rules
- Always check messages before starting work — someone may have warned about a conflict
- Announce significant file changes — other agents can't see your edits in real-time
- Reply to questions — don't leave messages unacknowledged
- Be concise — messages are for coordination, not essays
- Don't spam — one announcement per major action, not per file edit
