Ask HN: Why does MCP need to exist? Why can't we just use REST APIs with OpenAPI specs provided to the LLM (possibly alongside x-llm-* hints to give additional contextual information)? Arguments I've seen focus on advantages of MCP for stateful or bidirectional connections, but it's not obvious that additional complexity is worth the tradeoff to me. Help me understand why this exists |
There is quite the security risk it seems. From giving your credentials to access to your filesystem and other OS related stuff.
Would you go to an website and willingly give your credentials or filesystem access?
You don't really know what is happening in the middle.
Finally if you are building an AI wrapper, you are just adding more "wrappers" on top
- Small Language Models: While most current MCP servers are wrappers on APIs, MCP was designed to be more than that. Think of a SLM running on a local NPU using MCP to interface with the device itself - streaming real-time data between the hardware and the SLM.
- Cost: OpenAPI specs are huge, and including them in the context window for every request would add up.
[1] https://trevorloula.com/blog/what-is-mcp/#why-use-mcp-rather...
1. LLMs hallucinate and often forget to close a bracket or leave a field out. This still happens in JSON mode like Gemini when it's a feature.
2. JSON formatting uses a lot of unnecessary tokens. Comma, quotes, brackets, etc.
3. Extra tokens also mean extra "cognitive effort" for the LLMs. We changed to YAML from JSON and saw a 30% or so increase in output quality back with GPT-3.5.
4. The above can be fixed with more and more training, but why train for REST when you can build something better for it?
Just started reading about MCP the other day and I feel like I must be missing something because I don't see the advantage.