MCP Tools Explained
MCP tools are callable capabilities that a server exposes to an AI host. A tool has a name, a human-readable description, an input schema, and optionally an output schema. The host can discover the tool catalog and make selected tools available to a model.
Discovery and invocation
A client can list tools, then call a specific tool with structured arguments. In the 2026 specification, method and tool names are also represented in HTTP routing headers for Streamable HTTP, making tool traffic easier to route, meter, and authorize at gateways.
Tools should represent meaningful actions
A good tool describes a task an agent can reason about: create_invoice, search_orders, or deploy_service. A poor tool often leaks an implementation detail, requires too many low-level calls, or combines unrelated actions behind ambiguous parameters.
Tools are security boundaries
Calling a tool can read sensitive data or mutate production systems. Treat every exposed tool like an API endpoint reachable by a probabilistic planner. Use least-privilege credentials, narrow schemas, server-side authorization, and human approval for risky actions.