wechat-mcp-macos

An MCP (Model Context Protocol) server that lets AI assistants read, monitor, and send WeChat messages on macOS. Supports continuous conversation polling — your AI can watch group chats and auto-reply.

Works with Claude Code, Claude Desktop, and any MCP-compatible client.

一个 MCP 服务器，让 AI 助手在 macOS 上读取、监控和发送微信消息。支持持续对话轮询——让你的 AI 盯着群聊自动回复。

How It Works / 工作原理

Reading messages: Decrypts the local WeChat SQLite database (SQLCipher 4, AES-256-CBC) and queries it directly. This is fast, accurate, and doesn't require screenshots or OCR.

Sending messages: Uses macOS osascript (AppleScript) to automate the WeChat desktop client — sets the clipboard, clicks the input box, pastes, and hits Enter.

读取消息：解密本地微信 SQLite 数据库（SQLCipher 4, AES-256-CBC）并直接查询。速度快、准确率 100%，不需要截图或 OCR。

发送消息：通过 macOS osascript（AppleScript）自动化微信桌面客户端——设置剪贴板、点击输入框、粘贴、回车发送。

Requirements / 环境要求

macOS (tested on macOS 15 Sequoia)
WeChat desktop client (微信 Mac 版)
Python 3.10+
Xcode command-line tools (xcode-select --install)

Installation / 安装

# With uv (recommended)
uv tool install wechat-mcp-macos

# Or with pip
pip install wechat-mcp-macos

Connect to Claude Code / 连接 Claude Code

claude mcp add wechat -- wechat-mcp-macos

Connect to Claude Desktop / 连接 Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "wechat": {
      "command": "wechat-mcp-macos"
    }
  }
}

First-Time Setup / 首次配置

The first time you use it, you need to set up WeChat database access:

首次使用需要配置微信数据库访问：

1. Re-sign WeChat / 重签名微信

WeChat ships with hardened runtime, which prevents memory scanning for encryption keys. You need to re-sign it:

微信自带 hardened runtime，需要重签名才能提取密钥：

# Quit WeChat first! / 先退出微信！
# Then ask your AI to run:
setup("sign")

Or manually:

sudo codesign --force --deep --sign - /Applications/WeChat.app

Note: After re-signing, you may need to re-authorize WeChat's screen recording permission in System Settings → Privacy & Security → Screen Recording. Run tccutil reset ScreenCapture com.tencent.xinWeChat then restart WeChat.

注意：重签名后可能需要重新授权微信的录屏权限。

2. Extract encryption keys / 提取加密密钥

Start WeChat and log in, then:

启动微信并登录，然后：

setup("extract")

This compiles a C memory scanner, scans the WeChat process for SQLCipher keys, and saves them to ~/.wechat-mcp/all_keys.json.

这会编译一个 C 内存扫描器，从微信进程中提取 SQLCipher 密钥。

3. Verify / 验证

get_status()

If everything says "ready", you're good to go!

如果一切显示"ready"，就可以开始使用了！

MCP Tools / 可用工具

| Tool | Description | |------|-------------| | get_status() | Health check: config, keys, WeChat process status | | setup(step) | Guided setup: "detect", "sign", "extract", or "auto" | | list_chats(limit) | Recent conversations with timestamps and previews | | read_messages(chat, limit, since_hours) | Read messages from any chat | | check_new(chats, since_minutes) | Lightweight new-message check for polling | | search(keywords, chat_names, hours_back) | Cross-chat keyword search | | list_groups() | List all group chats | | send(text, chat) | Send a message (requires Accessibility permission) |

Conversation Monitoring / 对话轮询

The real power of this tool is continuous conversation monitoring. Instead of reading messages once, you can have Claude poll your chats and respond automatically — turning it into a 24/7 WeChat assistant.

这个工具真正强大的地方在于持续对话监控。不是读一次就完了，而是让 Claude 持续轮询你的聊天，自动回复——相当于一个 24 小时在线的微信助手。

How it works / 工作原理

check_new() is a lightweight tool that only returns message counts (no content), making it fast and cheap to call repeatedly
When new messages are detected, Claude reads them with read_messages()
Claude processes the content and optionally replies with send()

Just tell Claude what to do in natural language:

"Monitor the group chat '项目组' every 3 minutes.
 If someone asks a question, try to answer it.
 If someone @'s me, let me know."

"Watch my DMs with 妈妈. If she sends a message, reply with
 '我在忙，稍后回复' and summarize what she said."

"Poll these 3 groups every 5 minutes and give me a digest
 of anything important."

Claude will loop check_new() → read_messages() → send() on its own.

Claude 会自动循环调用 check_new() → read_messages() → send()。

Usage Examples / 使用示例

Once connected to your AI client:

"Show me my recent WeChat messages"
→ list_chats() + read_messages()

"Search for messages about 'dinner' in the last week"
→ search(keywords="dinner", hours_back=168)

"Send 'I'll be there in 10 minutes' to Mom"
→ send(text="I'll be there in 10 minutes", chat="Mom")

"Monitor '工作群' and auto-reply when someone asks about the deadline"
→ check_new() → read_messages() → send() (polling loop)

Sending Messages / 发送消息

To send messages, the app running the MCP server needs Accessibility permission:

发送消息需要辅助功能权限：

System Settings → Privacy & Security → Accessibility → Enable your terminal app / Claude.app

系统设置 → 隐私与安全性 → 辅助功能 → 启用终端 / Claude.app

Configuration / 配置

Config is stored at ~/.wechat-mcp/config.json:

{
  "db_dir": "/path/to/wechat/db_storage",
  "keys_file": "~/.wechat-mcp/all_keys.json",
  "decrypted_dir": "~/.wechat-mcp/decrypted",
  "self_name": "Me"
}

self_name: Display name for your own messages (they have empty sender in the DB)

If you previously used wechat-summary-share, config will be auto-migrated from ~/.wechat-summary/.

Known Limitations / 已知限制

macOS only — uses Mach VM APIs for key extraction and osascript for sending
Read-only database — the database is decrypted as a copy, original files are never modified
WeChat updates may break key extraction — re-sign and re-extract keys after updates
Cannot read image/video/voice content — only shows [图片] [视频] [语音] markers
Sending requires UI automation — WeChat window must be accessible

Credits / 致谢

ylytdeng/wechat-decrypt: C memory scanner and SQLCipher 4 decryption logic
Qizhan7/mac-wechat-summary: WeChatDB query interface, message parsing, FTS search, zstd decompression

License

MIT

MCP Servers