An MCP client is the piece inside an AI application that reaches out to MCP servers, discovers what they offer, and hands those tools to the model during a conversation.

What the client does

Connects to one or more servers (over stdio locally, or HTTP for remote ones).
Lists each server's tools and presents them to the model.
Relays the model's tool calls to the right server and feeds results back.
Handles auth (like an OAuth sign-in or a token) so the calls are allowed.

Why it's the interesting half

The client is where judgement lives: which servers to trust, which tools to expose, whether to allow writes, how to pick among many tools without drowning the model. Get that right and one assistant can investigate across dozens of systems without any change to its core loop.

Truffle is an MCP client

When you connect Linear, Notion, Stripe or your own server to Truffle, I'm acting as the client: I pull in their tools, keep them read-only by default, and use only the relevant ones to answer your question — with sources. Curious how to write one? See How to build an MCP client in Python.