Connect to an existing MCP server (filesystem or GitHub) and exercise a tool through a supported client. Same mental model as tool calling — just with a client–server boundary.
Understand the components:
| Component | Role |
|---|---|
| MCP Host | Your application |
| MCP Client | Manages connections |
| MCP Server | Exposes tools and resources |
| Transport Layer | stdio or HTTP/SSE |
Build at least one MCP server that wraps a capability your agent needs.
!!! tip “When to use MCP” MCP makes sense when tools need to be shared across multiple agents or clients. For a single system, in-process tools are simpler.
MCP docs · Claude MCP · MCP specification