WhatsApp MCP lets your AI search personal chats from your own Cloudflare account

Ask an AI agent what someone said about the flights, or find every mention of a lease across your WhatsApp chats. That is the workflow behind Eli Finer’s WhatsApp MCP, introduced on Show HN on October 18. The independent server runs in your own Cloudflare account and lets a compatible MCP host read and search personal messages. It cannot send messages.

The phone links through a QR code or phone-number pairing, much like adding a WhatsApp Web device. Claude Code, Claude Desktop or another HTTP MCP client then connects to the deployed /mcp endpoint. The host provides the AI model and its account; this is a connector, not a new model or an official WhatsApp integration.

Searchable history, with on-demand synchronization

The README describes one Durable Object with SQLite holding messages, chats, contact labels and the live WhatsApp login. The server connects to collect queued messages on demand and once daily at 04:00 UTC, then disconnects. Read tools refresh a stored copy older than ten minutes. This is not an always-connected real-time stream.

Initial history depends on what WhatsApp supplies to a newly linked device and may be incomplete. Photos, documents and voice notes appear as labels with available captions; the server neither downloads nor transcribes them. For unnamed contacts, the host agent can use its existing contacts connector and save matching names. The server itself does not log into Google Contacts.

The current hosting instructions qualify the free pitch

The Show HN title says the server runs free on Cloudflare. The current README instead requires Workers Paid, because importing a large history uses a raised CPU limit. The author says the free plan is untested and an initial import may be fragile after removing that limit. MIT-licensed code therefore does not establish free hosting or free model usage.

Installation requires Node.js 20 or later, cloning the repository, installing dependencies, creating an OAuth KV namespace, setting an owner password of at least 12 characters and deploying before pairing the phone. Each deployment is intended for one owner and one account.

Account access is material: the deployment holds message history and live login keys, and requested chat results pass to the chosen AI host. Self-hosting storage does not establish that model processing stays local. The README also says its unofficial Baileys connection violates WhatsApp’s terms. Its assessment of low practical risk is the author’s judgment, not a verified account-safety guarantee. AITrending has reviewed the documentation, not independently paired a personal account.