> The core philosophy of MCP is simple: it’s an API abstraction. The LLM doesn’t need to understand the how; it just needs to know the what.

Wrong. It needs to "understand" both these things. The only difference is where and how the strings explaining them are generated.