MCP — Connecting AI to Your Tools

How MCP Works

Hosts, Clients, and Servers

MCP splits the system into three clearly separated roles. Understanding these three is the core of understanding MCP.

Host — the application you interact with directly (like an AI desktop app or a coding IDE). It manages your session, enforces security boundaries, and decides which servers to connect to. One host can connect to many servers at once.

Client — lives inside the host and handles the actual protocol communication: discovering what a server offers, formatting messages, and parsing responses. Think of it as the translator.

Server — wraps a specific tool or data source (your database, your calendar, GitHub) and exposes what it can do in the MCP standard.

A Concrete Example

Picture an AI app (the host) with three servers connected at once: one for your database, one for your calendar, and one for your files. Each connection is handled by its own client inside the host. This separation creates clear trust boundaries — the calendar server can't reach into the database server's territory. The host stays in control of what's connected and what's allowed.

Host decides and secures, client translates, server does the actual work — three roles, cleanly separated

Hosts, Clients, and Servers — MCP — Connecting… | hivework