What it does well though is to keep things relatively simple. It is also good for things that takes time to return.
I find it competitive also for non LLM usages. I have done some PoCs with MCP instead of REST/OpenAPI, or GraphQL, or Protobuf and I liked MCP.
But yeah, a lot of it (including tools) maps well to HTTP semantics. In fact, at work we implement our features as HTTP APIs, and we’ve built a framework that exposes them as MCP servers.
If you're needing to give the model direct and broad access to your SqlServer instance...I don't get it.
Dedicated customer facing API is like whitelisting approved actions; MCP feels like getting an LLM to enforce a complex blacklist.
Now every website can create tools that gpt can use. So it's possible to have your users use AI without the need to pay for tokens.