Why it matters

Tools are how MCP servers enable LLM agents to take actions. Good tool design matters more than protocol details.

Advertisement

The architecture

list_tools: server returns tool definitions (name, description, input schema).

call_tool: client sends name + arguments; server executes; returns result.

MCP tool interactionTool definitionname + schema + descriptioncall_toolinvoke with argsResultstructured outputDescriptions drive LLM tool selection; schemas prevent invalid calls
Tool invocation flow.
Advertisement

How it works end to end

Descriptions: natural language for LLM tool selection. Be specific about when to use.

Schemas: JSON Schema for inputs. Enables validation.

Result format: structured content parts. Text, images, structured data.

Progress: for long-running tools, server can send progress notifications.