How does MCP work, step by step?

MCP works through a defined request flow. A host instructs a client; the client sends a JSON-RPC request over a transport to a server; the server handles it and returns a response; the client returns the result to the host. Initialization negotiates capabilities before any request is served.

First, the client initializes the connection and exchanges capabilities. Then the host asks the client for something a server can provide. The client sends the appropriate protocol request — tools/call, resources/read, or prompts/get — to the server. The server processes it and returns a structured response, which the client passes back to the host.

Most failures occur at the boundaries: the server is not running or not on the expected transport; capability negotiation failed so a requested primitive is unavailable; the transport dropped the connection; or the server rejected authentication. Because the protocol is clear, most problems are detectable by inspecting the client and server logs during initialization.