WhatsApp agent workflows

In the Gulf, business happens on WhatsApp. The messages arrive all day, and most of them ask for something a system already knows.

What was built. WhatsApp connected to an agent built on the Claude API. Incoming messages reach the agent through an MCP server that exposes WhatsApp as tools: read a conversation, send a reply, flag a thread for a person. A second set of MCP tools gives the agent scoped access to the business's own data, so it can answer from the record rather than from memory. Anything outside its scope is handed to a person with the context attached.

Stack. Claude API, Model Context Protocol servers in Node.js, the WhatsApp integration, with every tool call logged.

What was hard. Trust. The agent can only do what its tools allow, every action is written to a log the owner can read, and the first version was deliberately narrow: one kind of request, measured, then widened.