What is an MCP tool?

An MCP tool is one of the three core primitives a Model Context Protocol server can offer, and it is the action primitive: something the model can actively invoke to do work or fetch a result. Each tool has a name, a human-readable description, and a JSON Schema for its inputs, which together tell the model when and how to call it. At runtime the model decides to use a tool, the host asks the user to approve it (or applies a policy), the client sends a tools/call request, and the server executes and returns a structured result. Good tool design matters more than almost anything else in a server: descriptions should make the tool's purpose obvious, input schemas should be tight, and the surface should stay small so the model can choose well. Because tools are model-controlled and can have real side effects, like opening a pull request or writing to a database, they are the primitive that carries the most power and the most risk. Glen's memory server, for example, exposes a single glen tool that both recalls relevant context and records new observations in one round trip.