MS Labs internal tool, 2026
WhatsApp MCP connector for AI
An MCP server that gives AI assistants read-only access to WhatsApp Desktop conversations, reading the app's own local database and transcribing voice notes on the machine, with no sending and no external connections.
Stack: TypeScript, Model Context Protocol, SQLite, Whisper (local), ffmpeg
The problem
A lot of work context lives in WhatsApp conversations. Bringing it to an AI assistant usually means copy-pasting, or connecting third-party services that ask for full account access.
What we built
A local MCP (Model Context Protocol) server that lets an AI assistant read and search WhatsApp Desktop conversations, and transcribe voice notes, with three guarantees:
- Read-only: there is no sending tool at all;
- No external connection: it reads the database the app already keeps on disk, without creating a linked-device session;
- On-device transcription: voice notes are transcribed on the machine, never uploaded.
Why it matters
It's the same principle we apply to AI systems for clients: give the model the context it needs with the least power possible. A connector that can't send messages can't send the wrong message.
See how we build custom AI systems.