Letting your agent read your WhatsApp, and what to watch out for

WhatsApp is a hub for coordination, lacking an official API for personal accounts. When people inquire if an agent can summarize a family group's discussions, the straightforward response has two parts: yes, but with a catch. As someone working on LMCP, a free MCP server for Mac, I can explain how its WhatsApp tools operate. The first part is that LMCP connects to WhatsApp through wacli, an open-source unofficial client that acts as your device within WhatsApp.

This connection carries a warning in each WhatsApp tool's description, emphasizing the risk of violating WhatsApp's terms. If your agent links their device via whatsapp_connect, messages are stored locally on your Mac. LMCP treats these stored messages like credentials and refuses to read them. The agent's workflow involves finding the chat and its ID, reading messages (50 by default), and searching across chats.

Since there are no date filters, the agent summarizes what's been discussed. However, WhatsApp history arrives on the Mac gradually, with reads triggering a short background sync. If you're using automated syncing, there's a limit to how often it can occur, so aggressive automation can lead to being logged out. After linking your device, the first read might be a bit outdated, but subsequent reads will capture the latest updates.

If your phone shows more unread messages than the Mac has, the read result will indicate that the list is incomplete. Sending messages requires a confirmation step, and if the chat's name can't be resolved, the preview will alert you instead of making assumptions. For an agent that runs unattended, it's safer not to grant it access to send tools.

To try it out, you can use the command curl -fsSL https://local-mcp.com/install?ref=devto-t4 | bash.