🗓️ 19092026 2037

MCP ARCHITECTURE

MCP separates the AI application from capability providers. A host creates an MCP client for each MCP server it wants to use.

host
 ├─ client A <-> server A
 `─ client B <-> server B

The host owns the user experience and model integration. Each client maintains one connection. Servers provide capabilities over that connection.

What happens on a connection​

  • Client and server initialize and announce supported capabilities.
  • Client discovers available tools, resources, or prompts.
  • Client calls a tool or reads data when the host decides it is appropriate.
  • Either side closes the connection when work ends.

The initialization step lets both sides avoid using features the other side does not support.

Server and client capabilities​

Servers commonly provide tools, resources, and prompts. Clients can also offer services back to servers, such as asking the user for information or requesting a model completion. These are advanced cases; most integrations only need server tools.

For local versus remote connections, see mcp_transports. For the basic concepts, see the MCP overview.

References​