Definition
An MCP client is the protocol component an MCP host creates to establish and manage a connection with one MCP server. It performs initialization and capability negotiation, sends requests, receives responses and notifications, and exposes the server's available primitives to the host under local policy.
The architectural boundary is easy to miss because products often call the whole application an MCP client. In the protocol model, the host is the AI application that coordinates user interaction, models, security, and multiple server connections. It maintains a separate client connection for each server. That separation helps contain connection state and gives the host a place to apply consent and authorization decisions.
The model is not the client
A language model may choose to use a tool surfaced through MCP, but it does not normally speak the wire protocol or own the connection. The host translates between model-facing capability descriptions and protocol messages, then decides which results enter model context.
Distinguish it from nearby terms
The host is the application boundary. The client is the per-server protocol participant. The server exposes tools, resources, or prompts. SDK classes and product documentation sometimes use these names loosely, so architecture discussions should state which layer they mean.
Check your understanding
An AI desktop application connects to three MCP servers. How many protocol clients may it maintain? Typically three, one per server connection, even though the user sees one host application and one model conversation.