you have to write code for MCP server, and code to consume them too. It's just the LLM vendor decide that they are going to have the consume side built-in, which people question as they could as well do the same for open API, GRPC and what not, instead of a completely new thing.