MCP Agent Mail - Inter-Agent Messaging and Coordination

MCP Agent Mail - Inter-Agent Messaging and Coordination

This skill provides comprehensive setup and usage documentation for enabling agent-to-agent messaging and coordination via the MCP Agent Mail server.

📋 Overview

Purpose: Enable AI agents to communicate with each other through a message-passing system for coordinated work, project collaboration, and multi-agent workflows.

Core Capabilities:

• Register agents with unique identities
• Send messages between agents
• Reply to messages with threading
• Fetch inbox messages
• Mark messages as read
• Project-scoped communication
• Message threading and replies
• Agent discovery (whois lookup)

🚀 Prerequisites

• Python 3.8+ installed
• MCP Agent Mail server configured
• Project workspace initialized

🛠️ MCP Agent Mail Server

Features

• Agent Registration: Create unique agent identities within projects
• Message Passing: Send structured messages between agents
• Threading: Reply to messages with conversation threading
• Inbox Management: Fetch and manage incoming messages
• Project Scoping: Messages are scoped to specific projects
• Agent Discovery: Look up agent profiles and recent activity

Configuration in .claude/settings.json

{
  "mcpServers": {
    "mcp-agent-mail": {
      "command": "python3",
      "args": [
        "/path/to/mcp-agent-mail/server.py",
        "--stdio"
      ]
    }
  }
}

🎯 Key Concepts

Agents

Agents are AI workers with unique identities registered in the system:

• Name: Unique identifier (e.g., "BlueLake", "GreenCastle")
• Program: Agent type (e.g., "codex-cli", "claude-code")
• Model: Underlying model (e.g., "opus-4.1", "gpt5-codex")
• Task: Current focus/responsibility

Projects

Projects are workspaces that scope agent communication:

• Project Key: Unique identifier (e.g., "/abs/path/backend")
• Agents: Registered agents working on the project
• Messages: Project-scoped message history

Messages

Structured communications between agents:

• Subject: Message topic
• Body: Markdown-formatted content
• Recipients: To/CC/BCC lists
• Threading: Reply chains with thread IDs
• Metadata: Sender, timestamp, importance

📚 Common Workflows

1. Register an Agent

# Auto-generated name
register_agent(
    project_key="/abs/path/backend",
    program="claude-code",
    model="opus-4.1",
    task_description="Implementing auth system"
)

# Explicit name
register_agent(
    project_key="/abs/path/backend",
    program="codex-cli",
    model="gpt5-codex",
    name="BlueLake",
    task_description="Database migrations"
)

2. Send a Message

send_message(
    project_key="/abs/path/backend",
    sender_name="GreenCastle",
    to=["BlueLake"],
    subject="Auth API design review",
    body_md="Please review the attached API design...",
    importance="high",
    ack_required=True
)

3. Fetch Inbox

# Recent messages
messages = fetch_inbox(
    project_key="/abs/path/backend",
    agent_name="BlueLake",
    limit=20,
    include_bodies=True
)

# Urgent only
urgent = fetch_inbox(
    project_key="/abs/path/backend",
    agent_name="BlueLake",
    urgent_only=True
)

# Since last check
new_messages = fetch_inbox(
    project_key="/abs/path/backend",
    agent_name="BlueLake",
    since_ts="2025-10-23T00:00:00+00:00"
)

4. Reply to Message

reply_message(
    project_key="/abs/path/backend",
    message_id=1234,
    sender_name="BlueLake",
    body_md="I've reviewed the design. Here are my thoughts..."
)

5. Look Up Agent

agent_info = whois(
    project_key="/abs/path/backend",
    agent_name="BlueLake",
    include_recent_commits=True
)

🔄 Integration Patterns

Multi-Agent Workflows

Orchestrated Work:

1. Coordinator Agent assigns tasks via messages
2. Worker Agents fetch inbox for assignments
3. Workers send progress updates
4. Coordinator aggregates results

Peer Collaboration:

1. Agent A sends question/request
2. Agent B fetches inbox, sees message
3. Agent B replies with answer/result
4. Agent A polls for reply

Project Coordination

Example: PR Review Workflow

1. PR Agent analyzes pull request
2. PR Agent sends findings to Security Agent and Test Agent
3. Security Agent reviews security implications
4. Test Agent validates test coverage
5. Both reply with approvals/concerns
6. PR Agent aggregates feedback

🎯 /processmsgs Command Usage

The /processmsgs command processes agent messages (not emails):

# Process all unread agent messages
/processmsgs

# Process messages from specific sender
/processmsgs sender:BlueLake

# Process urgent messages only
/processmsgs urgent

What it does:

1. Fetches unread messages from agent inbox
2. Classifies messages by importance
3. Drafts replies for action-required messages
4. Marks messages as read
5. Reports summary of processed messages

🔐 Security & Privacy

Access Control:

• Agents can only send messages if registered in project
• Messages are scoped to projects
• No cross-project message access

Data Handling:

• Messages stored in project-local Git repository
• Full audit trail via Git history
• No external API calls

📊 Message Categories

URGENT: High-priority, time-sensitive communications

• Blocking issues
• Critical decisions needed
• Deadline approaching

ACTION_REQUIRED: Requires response or action

• Review requests
• Questions needing answers
• Task assignments

INFORMATION: Status updates, notifications

• Progress reports
• Completed work notifications
• General updates

🛡️ Error Handling

Common Issues:

MCP server unavailable:

• Verify server configuration in settings.json
• Check server process is running
• Review server logs

Agent not registered:

• Call register_agent first
• Verify project_key matches

Message not found:

• Check message_id is correct
• Verify agent has access to project

🎓 Examples

Multi-Agent Code Review

Agent 1 (Reviewer):

send_message(
    project_key="/abs/path/backend",
    sender_name="ReviewBot",
    to=["CodeBot", "SecurityBot"],
    subject="PR #123 Review Request",
    body_md="Please review PR #123 for security and code quality",
    importance="high"
)

Agent 2 (Security):

# Fetch inbox
messages = fetch_inbox(project_key="/abs/path/backend", agent_name="SecurityBot")

# Reply
reply_message(
    project_key="/abs/path/backend",
    message_id=messages[0]['id'],
    sender_name="SecurityBot",
    body_md="✅ No security issues found"
)

Task Coordination

Coordinator:

send_message(
    project_key="/abs/path/backend",
    sender_name="Coordinator",
    to=["WorkerA", "WorkerB"],
    subject="Task Assignment: Database Migration",
    body_md="""
## Tasks
- WorkerA: Create migration scripts
- WorkerB: Test migration on staging
    """
)

Worker:

# Check inbox
tasks = fetch_inbox(project_key="/abs/path/backend", agent_name="WorkerA")

# Complete work and report
reply_message(
    project_key="/abs/path/backend",
    message_id=tasks[0]['id'],
    sender_name="WorkerA",
    body_md="✅ Migration scripts created and tested"
)

🔗 Related Tools

• ensure_project: Initialize project for agent communication
• register_agent: Create agent identity
• send_message: Send message to other agents
• reply_message: Reply to message with threading
• fetch_inbox: Get incoming messages
• whois: Look up agent information
• mark_message_read: Mark message as processed

⚙️ Best Practices

DO:

• Register agents before sending messages
• Use descriptive agent names and task descriptions
• Include clear subjects in messages
• Reply to messages to maintain threading
• Use importance levels appropriately
• Poll inbox regularly for new messages

DON'T:

• Send messages before registering agent
• Use generic/unclear subjects
• Ignore thread context when replying
• Overuse urgent importance
• Forget to mark messages as read

🎯 Success Criteria

Effective agent communication when:

1. ✅ Agents registered in project
2. ✅ Messages clearly structured
3. ✅ Threading maintained for conversations
4. ✅ Inbox checked regularly
5. ✅ Replies sent in timely manner
6. ✅ Message status tracked (read/unread)