WhatsApp MCP Server (TypeScript/Baileys)

This is a Model Context Protocol (MCP) server for WhatsApp, built with TypeScript and using the @whiskeysockets/baileys library.

It allows you to connect your personal WhatsApp account to an AI agent (like Anthropic Claude via its desktop app or Cursor) enabling it to:

• Search your personal WhatsApp messages.
• Search your contacts (individuals, not groups).
• List your recent chats.
• Retrieve message history for specific chats.
• Send messages to individuals or groups.

It connects directly to your personal WhatsApp account using the WhatsApp Web multi-device API. All your messages and authentication details are stored locally in a SQLite database (./data/) and authentication cache (./auth_info/). Data is only sent to the connected AI agent when it explicitly uses the provided MCP tools (which you control via the agent's interface).

(Optional: Consider adding a screenshot or GIF similar to the reference example here)

Example

User: Send a whatsapp message to "Meu amor" in whatsapp saying "Te amo"

Assistant: Okay, I need to find the contact first. Using tool: whatsapp.search_contacts

{
  "query": "Meu amor"
}

Tool Result:

[
  {
    "jid": "5599xxxxxx@s.whatsapp.net",
    "name": "Meu Amor"
  }
]

Assistant: Found the contact. Now sending the message. Using tool: whatsapp.send_message

{
  "recipient": "5599xxxxxx@s.whatsapp.net",
  "message": "Te amo"
}

Tool Result:

Message sent successfully to 5599xxxxxx@s.whatsapp.net (ID: XXXXXXXXXXX).

Key Features (MCP Tools)

The server exposes the following tools to the connected AI agent:

• search_contacts: Search for contacts by name or phone number part (JID).
• list_messages: Retrieve message history for a specific chat, with pagination.
• list_chats: List your chats, sortable by activity or name, filterable, paginated, optionally includes last message details.
• get_chat: Get detailed information about a specific chat.
• get_message_context: Retrieve messages sent immediately before and after a specific message ID for context.
• send_message: Send a text message to a specified recipient JID (user or group).

Installation

Prerequisites

• Node.js: Version 23.10.0 or higher (as specified in package.json). You can check your version with node -v. (Has initial typescript and sqlite builtin support)
• npm (or yarn/pnpm): Usually comes with Node.js.
• AI Client: Anthropic Claude Desktop app, Cursor, Cline or Roo Code (or another MCP-compatible client).

Steps

1. Clone this repository:

   git clone <your-repo-url> whatsapp-mcp-ts
   cd whatsapp-mcp-ts
   

2. Install dependencies:

   npm install
   # or yarn install / pnpm install
   

3. Run the server for the first time: Use node to run the main script directly.

   node src/main.ts
   

   • The first time you run it, it will likely generate a QR code link using quickchart.io and attempt to open it in your default browser.
   • Scan this QR code using your WhatsApp mobile app (Settings > Linked Devices > Link a Device).
   • Authentication credentials will be saved locally in the auth_info/ directory (this is ignored by git).
   • Messages will start syncing and be stored in ./data/whatsapp.db. This might take some time depending on your history size. Check the wa-logs.txt and console output for progress.
   • Keep this terminal window running. After syncing you can close.

Configuration for AI Client

You need to tell your AI client how to start this MCP server.

1. Prepare the configuration JSON: Copy the following JSON structure. You'll need to replace {{PATH_TO_REPO}} with the absolute path to the directory where you cloned this repository.

   {
     "mcpServers": {
       "whatsapp": {
         "command": "node",
         "args": [
           "{{PATH_TO_REPO}}/src/main.ts"
         ],
         "timeout": 15, // Optional: Adjust startup timeout if needed
         "disabled": false
       }
     }
   }
   

   • Get the absolute path: Navigate to the whatsapp-mcp-ts directory in your terminal and run pwd. Use this output for {{PATH_TO_REPO}}.

2. Save the configuration file:

   • For Claude Desktop: Save the JSON as claude_desktop_config.json in its configuration directory:
     • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
     • Windows: %APPDATA%\Claude\claude_desktop_config.json (Likely path, verify if needed)
     • Linux: ~/.config/Claude/claude_desktop_config.json (Likely path, verify if needed)
   • For Cursor: Save the JSON as mcp.json in its configuration directory:
     • ~/.cursor/mcp.json

3. Restart Claude Desktop / Cursor: Close and reopen your AI client. It should now detect the "whatsapp" MCP server and allow you to use its tools.

Usage

Once the server is running (either manually via node src/main.ts or started by the AI client via the config file) and connected to your AI client, you can interact with your WhatsApp data through the agent's chat interface. Ask it to search contacts, list recent chats, read messages, or send messages.

Architecture Overview

This application is a single Node.js process that:

1. Uses @whiskeysockets/baileys to connect to the WhatsApp Web API, handling authentication and real-time events.
2. Stores WhatsApp chats and messages locally in a SQLite database (./data/whatsapp.db) using node:sqlite.
3. Runs an MCP server using @modelcontextprotocol/sdk that listens for requests from an AI client over standard input/output (stdio).
4. Provides MCP tools that query the local SQLite database or use the Baileys socket to send messages.
5. Uses pino for logging activity (wa-logs.txt for WhatsApp events, mcp-logs.txt for MCP server activity).

Data Storage & Privacy

• Authentication: Your WhatsApp connection credentials are stored locally in the ./auth_info/ directory.
• Messages & Chats: Your message history and chat metadata are stored locally in the ./data/whatsapp.db SQLite file.
• Local Data: Both auth_info/ and data/ are included in .gitignore to prevent accidental commits. Treat these directories as sensitive.
• LLM Interaction: Data is only sent to the connected Large Language Model (LLM) when the AI agent actively uses one of the provided MCP tools (e.g., list_messages, send_message). The server itself does not proactively send your data anywhere else.

Technical Details

• Language: TypeScript
• Runtime: Node.js (>= v23.10.0)
• WhatsApp API: @whiskeysockets/baileys
• MCP SDK: @modelcontextprotocol/sdk
• Database: node:sqlite (Bundled SQLite)
• Logging: pino
• Schema Validation: zod (for MCP tool inputs)

Troubleshooting

• QR Code Issues:
  • If the QR code link doesn't open automatically, check the console output for the quickchart.io URL and open it manually.
  • Ensure you scan the QR code promptly with your phone's WhatsApp app.
• Authentication Failures / Logged Out:
  • If the connection closes with a DisconnectReason.loggedOut error, you need to re-authenticate. Stop the server, delete the ./auth_info/ directory, and restart the server (node src/main.ts) to get a new QR code.
• Message Sync Issues:
  • Initial sync can take time. Check wa-logs.txt for activity.
  • If messages seem out of sync or missing, you might need a full reset. Stop the server, delete both ./auth_info/ and ./data/ directories, then restart the server to re-authenticate and resync history.
• MCP Connection Problems (Claude/Cursor):
  • Double-check the command and args (especially the {{PATH_TO_REPO}}) in your claude_desktop_config.json or mcp.json. Ensure the path is absolute and correct.
  • Verify Node.js are correctly installed and in your system's PATH.
  • Check the AI client's logs for errors related to starting the MCP server.
  • Check this server's logs (mcp-logs.txt) for MCP-related errors.
• Errors Sending Messages:
  • Ensure the recipient JID is correct (e.g., number@s.whatsapp.net for users, groupid@g.us for groups).
  • Check wa-logs.txt for specific errors from Baileys.
• General Issues: Check both wa-logs.txt and mcp-logs.txt for detailed error messages.

For further MCP integration issues, refer to the official MCP documentation.

Credits

• https://github.com/lharries/whatsapp-mcp Do the same as this codebase but uses go and python.

License

This project is licensed under the ISC License (see package.json).