Claude Desktop, Cursor, and other AI tools already support MCP. You configure some servers, and the tools appear. So why would you ever need to build your own MCP client?

Because the moment you need product-specific behavior, a pre-built host becomes limiting. You may want a Slack bot that uses approved MCP servers, a scheduled pipeline that reads MCP resources, a custom agent runtime with its own planning policy, or a multi-server tool router that exposes only a safe subset of available tools.

Pre-built hosts are excellent for interactive use. They are not where you put application-specific authorization, retry policy, audit logging, tool filtering, or workflow orchestration. For that, you build the host/client layer yourself.

The Python SDK gives you the protocol machinery. Your job is to build the product machinery around it: connection lifecycle, capability discovery, model integration, routing, error handling, and policy.