i really dont get why we cant just feed the openapi spec to the LLM instead of having this intermediate MCP representation. Don't really buy the whole 'the api docs will overwhelm an LLM" - that hasn't been my experience.
Really trivial to have the LLM first filter it down to the sections it cares about and then condense those sections though.
Wrap that process in a small tool and give that to the LLM along with a `fetch` tool that handles credentials based on URLs and agent capabilities explode pretty rapidly.
I agree, I wish, it will be a solved problem eventually. Just feeding a complex data model like that to the paper shredder that is the LLM, for making decisions about whether DELETE or POST is used is just asking for trouble.
Even if you're LLM could learn the openai spec, you still have to figure out how to concretely receive a response back. This is necessary for virtually any application build using an LLM and requires support for far, far more use cases than just calling an API.
Consider the following use case: - You need to include some relevant contextual data from a local RAG system. - There are local functions that you want the model to be able to call - The API example you describe - You need to access data from a database
In all of these cases, if you have experience working with LLMs, you've implemented some ad hoc template solution to pass the context into the model. You might have writing something like "Here is the info relevant to this task {{info}}" or "These are the tools you can use {{tools}}", but in each case you've had to craft a prompting solution specific to one problem.
MCP solves this by making a generic interface to sending a wide range of information to the model to make use of. While the hype can be a bit much, it's a pretty good (minus the lack of foresight around security) and obvious solution to this current problem in AI Engineering.
You don't need a spec.
For sending prompts to the LLM you will absolutely need to hand-craft custom prompts anyways, as each model responds slightly different.
Isn't that handled by whatever Tool API you're using? There's usually a `function_call_output` or `tool_result` message type. I haven't had a need for a separate protocol just to send responses.