Google MailPilot

Google MailPilot

Enables AI agents to read, triage, respond to, and calendar-manage Gmail with a production-grade MCP control plane and human-in-the-loop safety.

Category
访问服务器

README

Google MailPilot

Version Status License

The AI-native command center for Gmail and Google Workspace

Google MailPilot is the renamed successor to Gmail Secretary, built so LLMs can read, triage, respond to, and calendar-manage Gmail with a production-grade MCP control plane.

⚠️ Release Notice: v5.0.0 is the first public release under the Google MailPilot brand; it fixes MCP tool registration, tightens triage/compose coverage, adds new booking-link helpers, and keeps every mutation human-in-the-loop.

Release highlights:

  • Tools now register as soon as workspace_secretary.tools imports, preventing FastMCP registration bugs.
  • Triage/compose tooling gained synchronous tests plus tests/test_triage_priority_emails.py and tests/test_web_compose.py to cover CLI-aware workflows.
  • Engine/web routes, booking links, and raw: continuation-state handling now align with Google MailPilot expectations.

🚦 Stability Matrix

We believe in transparency. Here is the current readiness of our stack:

Component Status Stability Notes
IMAP Sync Engine 🟢 Stable RFC-compliant, IDLE support, tested on >24k emails.
Read Operations 🟢 Stable Search, Threading (X-GM-THRID), and Summarization work perfectly.
Write Operations 🟡 Beta Draft creation is stable. Auto-sending is currently disabled via the Safety Interceptor.
Web Dashboard 🟠 Alpha Early preview. UI may change rapidly.

A Google MailPilot IMAP/SMTP control plane for AI agents with calendar integration

Built for LLMs that need to read, search, triage, and respond to email autonomously. Not just an MCP wrapper — a full-featured IMAP control plane engineered for AI orchestration workflows.

📚 Documentation · 🏗️ Architecture · Quick Start


Why This Exists

Most email integrations for AI are thin API wrappers. They poll. They re-fetch. They timeout. They don't understand email threading, modification sequences, or push notifications.

Google MailPilot is different. It's a production-grade IMAP control plane that:

  • Syncs intelligently — CONDSTORE tracks what changed, IDLE pushes new mail instantly
  • Caches locally — SQLite or PostgreSQL, your AI reads from local DB in milliseconds
  • Understands Gmail — Native X-GM-THRID threading, X-GM-LABELS, X-GM-RAW search
  • Never sends without approval — Human-in-the-loop by design, drafts first

📡 RFC Compliance

We implement these IMAP extensions for efficient, real-time email sync:

RFC Extension What It Does Benefit
RFC 3501 IMAP4rev1 Core protocol Full IMAP compliance
RFC 7162 CONDSTORE Tracks modification sequences Skip sync when mailbox unchanged
RFC 7162 CHANGEDSINCE Fetch only changed flags 10x faster incremental sync
RFC 2177 IDLE Push notifications Instant new mail detection
RFC 2971 ID Client identification Better server compatibility

Gmail-Specific Extensions

Extension Purpose
X-GM-THRID Native Gmail thread ID — no heuristic threading needed
X-GM-MSGID Stable message identifier across folders
X-GM-LABELS Full Gmail label support (stored as JSONB)
X-GM-RAW Gmail's powerful search syntax for targeted sync

⚡ Performance

Real benchmarks against a 50,000 email mailbox:

Operation Traditional IMAP Google MailPilot
Check for new mail 2-5s (fetch all UIDs) < 50ms (HIGHESTMODSEQ compare)
Sync flag changes Re-fetch messages Flags only (CHANGEDSINCE)
New mail notification 5-min poll interval Instant (IDLE push)
Search emails Server roundtrip < 10ms (local SQLite FTS5)
Thread reconstruction Parse References headers Instant (X-GM-THRID)

How CONDSTORE Works

Traditional Sync:
  1. SELECT INBOX
  2. FETCH 1:* (FLAGS)        ← Downloads ALL flags every time
  3. Compare with cache
  4. Fetch changed messages

CONDSTORE Sync:
  1. SELECT INBOX
  2. Compare HIGHESTMODSEQ    ← Single integer comparison
  3. If unchanged → done      ← Skip everything
  4. If changed → FETCH 1:* (FLAGS) CHANGEDSINCE <modseq>
                              ← Only changed messages

🏗️ Architecture

Dual-process design separating sync from AI interface:

┌─────────────────────────────────────────────────────────────────┐
│                        AI Layer (Claude, etc.)                  │
└─────────────────────────────────────────────────────────────────┘
                                │
                                ▼
┌─────────────────────────────────────────────────────────────────┐
│                     MCP Server (read-only)                      │
│  • Exposes 25+ tools for email/calendar operations              │
│  • Reads directly from local database                           │
│  • Mutations proxied to Engine API                              │
└─────────────────────────────────────────────────────────────────┘
                                │
              ┌─────────────────┴─────────────────┐
              ▼                                   ▼
┌──────────────────────────┐        ┌──────────────────────────┐
│   SQLite / PostgreSQL    │        │     Secretary Engine     │
│  • Email cache (FTS5)    │◀───────│  • IMAP sync (CONDSTORE) │
│  • Gmail labels (JSONB)  │        │  • IDLE monitor          │
│  • Embeddings (pgvector) │        │  • OAuth2 management     │
└──────────────────────────┘        │  • SMTP send/draft       │
                                    └──────────────────────────┘
                                              │
                                              ▼
                                    ┌──────────────────┐
                                    │   Gmail IMAP     │
                                    │   Gmail SMTP     │
                                    │   Calendar API   │
                                    └──────────────────┘

Why Two Processes?

Concern Engine MCP Server
Sync Owns IMAP connection, IDLE loop Never touches IMAP
Database All writes Read-only
Uptime Always running Scales with AI requests
Credentials Holds OAuth tokens Stateless

🎯 AI-Native Features

Calendar Integration:

  • Instant Calendar Access: Sub-50ms queries via intelligent caching layer
  • 🌐 Offline-First Operations: Create/edit/delete events without internet, sync transparently
  • 🔄 Background Sync Worker: Autonomous daemon syncs every 60s using Google Calendar API sync tokens
  • 🏷️ Status Indicators: Visual badges show pending/synced/conflict states for offline operations
  • ⚙️ Multi-Calendar Support: Select which calendars to display via web UI settings

Email Intelligence:

Intelligent Signal Extraction

Every email is analyzed for actionable signals:

signals = {
    "is_addressed_to_me": True,      # In To: field, not just CC
    "mentions_my_name": True,        # Name appears in body
    "has_question": True,            # Contains "?" or question phrases
    "mentions_deadline": True,       # "EOD", "ASAP", "by Friday"
    "mentions_meeting": True,        # Scheduling keywords detected
    "is_from_vip": True,             # Sender in configured VIP list
    "has_attachments": True,         # PDF, DOCX, etc.
    "attachment_filenames": ["Q4_Report.pdf"]
}

Human-in-the-Loop Safety

The AI never sends email without explicit approval.

User: "Reply to Sarah saying I'll attend the meeting"

AI: I've drafted this reply:

    To: sarah@company.com
    Subject: Re: Team Meeting Tomorrow

    Hi Sarah,

    I'll be there. Looking forward to it!

    Best regards

    Send this email? (yes/no)

User: "yes"
AI: ✓ Email sent successfully

Confidence-Based Batch Operations

Bulk operations require approval with confidence tiers:

Confidence Batch Size Display
High (>90%) Up to 100 Date, From, Subject only
Medium (50-90%) Up to 10 + First 300 chars of body
Low (<50%) Individual Full context required

🚀 Quick Start

Prerequisites

  • Docker and Docker Compose
  • Gmail account with App Password (or OAuth2)
  • Google Cloud project (for Calendar integration)

1. Clone and Configure

git clone https://github.com/johnneerdael/google-mailpilot.git
cd google-mailpilot
cp config.sample.yaml config/config.yaml

2. Edit Configuration

# config/config.yaml
imap:
  host: imap.gmail.com
  port: 993
  username: your-email@gmail.com
  password: your-app-password    # Gmail App Password

user:
  email: your-email@gmail.com
  first_name: Your
  last_name: Name
  timezone: America/New_York
  working_hours:
    start: "09:00"
    end: "18:00"
  vip_senders:
    - ceo@company.com
    - important-client@example.com

database:
  backend: sqlite               # or "postgres" for embeddings
  path: /app/config/secretary.db

bearer_auth:
  enabled: true
  token: "generate-with-uuidgen"

3. Start Services

docker-compose up -d

4. Connect Your AI

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "secretary": {
      "url": "http://localhost:8000/mcp",
      "headers": {
        "Authorization": "Bearer your-token-here"
      }
    }
  }
}

🔧 Available Tools

Email Operations

Tool Description
search_emails FTS5-powered local search with Gmail query syntax
get_email_details Full email content with signals and metadata
get_email_thread Complete thread via X-GM-THRID
summarize_thread AI-ready thread summary
create_draft_reply Draft response (never auto-sends)
send_email Send with explicit approval
modify_gmail_labels Add/remove Gmail labels
move_email Move between folders

Calendar Operations

Tool Description
list_calendar_events Events in time range
get_calendar_availability Free/busy lookup
create_calendar_event Create with timezone support
suggest_reschedule Find alternative meeting times

Triage Operations

Tool Description
get_daily_briefing Priority emails + today's calendar
triage_priority_emails Identify high-priority items
quick_clean_inbox Batch cleanup with approval

Semantic Search (PostgreSQL + pgvector)

Tool Description
semantic_search_emails Search by meaning, not keywords
find_related_emails Similar emails to reference

📊 Database Schema

Email Storage

CREATE TABLE emails (
    uid INTEGER,
    folder TEXT,
    message_id TEXT UNIQUE,
    gmail_thread_id BIGINT,        -- X-GM-THRID
    gmail_msgid BIGINT,            -- X-GM-MSGID
    gmail_labels JSONB,            -- Full label set
    subject TEXT,
    from_addr TEXT,
    to_addr TEXT,
    date TIMESTAMPTZ,
    internal_date TIMESTAMPTZ,     -- INTERNALDATE
    body_text TEXT,
    body_html TEXT,
    flags TEXT,
    modseq BIGINT,                 -- CONDSTORE sequence
    has_attachments BOOLEAN,
    attachment_filenames JSONB,
    -- FTS5 index on subject, from_addr, to_addr, body_text
);

Folder State (CONDSTORE)

CREATE TABLE folder_state (
    folder TEXT PRIMARY KEY,
    uidvalidity INTEGER,
    uidnext INTEGER,
    highestmodseq BIGINT           -- For CONDSTORE sync
);

🔒 Security

Layer Protection
Transport TLS 1.2+ for IMAP/SMTP
Authentication OAuth2 or App Passwords (never plain passwords)
API Bearer token authentication
Data Local database, no cloud sync
Actions Human approval for all mutations

Never Stored

  • Plain text passwords
  • OAuth refresh tokens in logs
  • Email content in error messages

📚 Documentation

Guide Description
Architecture Deep dive into dual-process design
Configuration All config options explained
Agent Rules HITL safety patterns
API Reference Complete tool documentation

🛠️ Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run locally
python -m workspace_secretary.engine.api &  # Start engine
python -m workspace_secretary.server        # Start MCP server

License

MIT License — see LICENSE for details.


<p align="center"> <strong>Built for AI agents that take email seriously.</strong><br> <a href="https://github.com/johnneerdael/google-mailpilot">GitHub</a> · <a href="https://johnneerdael.github.io/google-mailpilot/">Documentation</a> · <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> </p>

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选