What is an MCP client?

MCP client, explained#

People often say "MCP client" to mean the whole AI app, and in conversation that is fine. The specification is more precise. The host is the application: it owns the user interface, the model and the consent decisions. The host creates one client per server it connects to, and each client keeps a one-to-one session with its server.

That split matters for security. Each client only sees its own server, so a server cannot read another server's data through the host. The host decides which tools the model may call and asks the user before running them; the client carries out the protocol.

A client's job, in order: connect over stdio or Streamable HTTP; send initialize with the protocol version and its own capabilities; read the server's capabilities; list tools, resources and prompts; and then, as the conversation goes on, call tools/call when the model picks a tool, fetch resources the user attaches, and handle notifications such as a changed tool list. For a remote server it also handles OAuth: following the 401 challenge, registering, opening the browser, storing and refreshing tokens, and re-authorizing when a call needs more scope.

Clients can offer features back to servers. Sampling lets a server ask the host's model for a completion, roots tell a server which folders it may touch, and elicitation lets a server ask the user a structured question mid-task.

For example, in Claude Code you run claude mcp add --transport http tracker https://example.com/mcp. Claude Code is the host; the entry you added becomes a client. When you type /mcp and choose Authenticate, that client performs the OAuth flow, and its tools appear for the model to use.