I've reread the article a couple times and I can't see where it says that you need to use an LLM to make it universal. I'm going to go so far as to say that it doesn't say that.
1. No one asserted that MCP's key innovation was its format. In fact, what I strongly, almost explicitly implied with my previous comment is that MCP has made no key innovations beyond its marketing system (which does matter and has forced a lot of companies who would never have operated in this space to make an MCP system, and that is good. I like MCP.).
2. MCP is fundamentally bidirectional, not unidirectional. Its a schematized request-response protocol that defines a ton of the same primitives that OpenAPI defines, like available RPC methods, authorization, request parameters, and response schemas [1]. Of course, OpenAPI goes beyond that into HTTP-specific characteristics like status codes, and MCP goes beyond that into LLM-specific characteristics like a prompt catalogue.
3. I'm not aware of any problem domain outside of LLMs that is adopting MCP. In fact, its adoption even within the world of LLMs has been lackluster at best [2].
[1] https://modelcontextprotocol.io/docs/learn/architecture#data...
And why? Because it dumbs it down to the point that the LLM can understand it (often, in part, by removing security concerns). But these are LLMs! They're supposed to be smart! Isn't it a supposedly temporary failure that I can't just point the LLM at an OpenAPI spec (or use HATEOAS) and be good to go?
Will this be doable in the next few months with better models? If so, why bother with MCP? If this won't be doable in the next few months / years, then how smart do we really expect these LLMs to be?
The reason why MCP was invented is hard to reverse-engineer, but from what I've seen I suspect its mostly: it was stdio/stdout first, and there isn't really a great standard like OpenAPI in the local-first stdio/stdout world. Interesting, LSP (Microsoft) took the opposite approach; despite having far less reason to be hosted on the web, LSP is just localhost http calls. MCP could have done that, but they didn't, without good reason, which leads me to the second reason why I suspect it was invented: they didn't know, and thus couldn't learn from prior art that they didn't know. Most of these AI labs are not rigorous engineering shops, they're better characterized as: a bunch of kids with infinite money trying to reignite the spirit of the early internet startup boom. Many believe that what they're building will replace everything, so why even try to build on what already exists?
But why it was invented doesn't matter, because anyone can invent anything. The reason why MCP has gotten popular is because its AI. That's it. Its a marketing and SEO system that enables companies who are otherwise struggling to find a way to prioritize AI use-cases for their apps to say "look, we're AI now". You don't get the same market impact if all you need is that boring old OpenAPI schema you already have.
And again: I like MCP. What I stated sounds pessimistic, but its only intended to be realistic. There is almost zero technical reason for MCP to exist beyond that.
This is neither an accurate quote of the article, nor an accurate paraphrase of something the article says, nor an accurate description of something subtly implied by the article without directly being said.
The article, in fact, gives an example without AI in the middle.
MCP has very little utility outside of LLMs. The article begins by saying "but what if we remove the AI", then goes back on that thesis by saying "but if there was an AI between these things, then it becomes a universal plugin system". Which is true, but its missing the critical bit: The AI itself is the universal plugin system, not MCP. And, beyond that, its not even the AI: Its natural language. Language is the universal plugin system.
Its not unbelievable that there exists an alternate reality where the Anthropic researchers that invented MCP instead simply leveraged a slightly extended form of OpenAPI specs. The only functional difference is that MCP was a stdin/stdout format first, and added HTTPS later, but sister problems in this domain like LSP just skipped stdin/stdout and went straight to locally-hosted HTTPS anyway. What matters isn't MCP, OpenAPI, or anything like that; what matters is the LLM itself, its tool calling capability, and the tool calling harness. The catalogue of available tools can be in any format; and, truly, the LLM does not care what format its in.