MCP describes the general protocol connecting models to tools and data. The MCP client is the side of that connection that lives inside the application or agent — it's what actually speaks the protocol to discover what an MCP server offers, and to call its tools or read its resources on the agent's behalf.

• Discovery — asking a connected server what tools and resources it exposes, so the agent knows what's available without hardcoding that knowledge in advance.
• Invocation — translating the agent's decision to use a tool into the actual protocol messages the server expects, and returning the result.
• Connection management — an agent might connect to several MCP servers at once, and the client manages each of those connections.
• Presenting capabilities to the model — the client typically surfaces discovered tools to the underlying model in the same way any other tool would be presented, so the model can decide when to use them.

• Assuming the client and the model are the same thing

  The client handles the protocol mechanics of discovering and calling tools; the model still decides which tool to use and with what arguments.

• Connecting to an MCP server without considering trust

  A server's tools and resources should be treated with the same scrutiny as any other tool source — an untrusted server is a real risk surface.

• Not handling a server becoming unavailable

  Like any other network dependency, an MCP server can be slow or go down — the client and the surrounding agent need a defined behavior for that case.