MCP Client

An MCP client is the "consumer" side of an MCP connection. It lives inside a host application — a coding agent, a chat assistant, an IDE plugin — and is responsible for opening connections to MCP servers, asking each one what it supports, and passing that information up to the model or the host's own logic so it can decide what to use.

A single host application commonly runs multiple MCP client connections at once, one per server it's configured to use. A coding agent might simultaneously connect to a filesystem server, a version-control server, and a project-management server, each through its own client connection, while presenting the model with a combined view of everything available.

The client doesn't decide on its own what the model is allowed to do — that policy typically lives in the host application, which might filter which servers are connected, require approval before certain tool calls run, or limit which resources get surfaced.

How it works

When a host starts up, its MCP client(s) connect to each configured server and perform capability negotiation, receiving back that server's list of tools, resources, and prompt templates. The client then exposes those to the model as available options — usually folded into the same interface the model already uses for calling tools. When the model decides to use one, the client sends the corresponding JSON-RPC request to the right server and returns the server's response. Because each server connection is independent, one server being slow or unavailable doesn't have to block the others.

Example

A chat application's MCP client might hold three simultaneous connections — to a search server, a calendar server, and a note-taking server — and merge all three servers' tools into one list the model can choose from during a conversation.

Common misconceptions

FAQ