What is MCP Client?
Also called Model Context Protocol client, MCP host.
An MCP client is the component inside an AI application that maintains a connection to a single MCP server and relays messages between that server and the application. The application that creates one or more clients is usually called the host. Clients handle version and capability negotiation, request routing, and returning tool results to the model.
When a host starts, it creates one client per configured server. Each client performs an initialization exchange that agrees on a protocol version and declares what each side supports. The client then lists the server's tools, resources, and prompts, and passes those descriptions into the model's context so the model can choose among them during a conversation.
Clients can also offer capabilities of their own. Some implementations let a server ask the client to run a model completion on its behalf, ask the user a question, or learn which directories it is permitted to work in. Support for these is optional and uneven across implementations, so a server that depends on them may not work everywhere it is installed.
People often conflate client and host. The host is the product a person actually uses, such as a chat application or a coding tool, while the client is the plumbing object the host creates per server. The distinction matters because approval prompts, credential storage, and audit logging are host responsibilities rather than protocol features.
Because the client is the point where a model's chosen call becomes a real action, it is the natural place for human approval gates, allow lists, and logging. A client that executes every requested tool call without any policy inherits every risk carried by the servers it is connected to.
Key points
- Lives inside the host application, one client per connected server.
- Negotiates protocol version and capabilities during initialization.
- Lists server tools and passes their descriptions into model context.
- Natural place for approval prompts, allow lists, and audit logs.
- The host is the user facing product; the client is per-server plumbing.
In practice
A desktop chat application is configured with two servers, one for a file system and one for a calendar. On launch it creates two clients, each negotiating separately. The user asks about tomorrow's schedule, the model chooses a calendar tool, and the calendar client sends the request, waits for the response, and hands the events back to the model. The file system client stays idle throughout.