The idea of MCP: “Tea. Earl grey. Hot.”

The article A Critical Look at MCP by Rasmus Holm (Raz Blog) discussing many of the drawbacks of MCP as it is currently conceived. One of them is the push to build everything in Python, which is a dynamic language that’s better-designed than JavaScript, but isn’t a lot better at helping users write maintainable code.

“Am I being pretentious/judgmental in thinking that people in AI only really know Python, and the “well, it works on my computer” approach is still considered acceptable? This should be glaringly obvious to anyone that ever tried to run anything from Hugging Face.

“If you want to run MCP locally, wouldn’t you prefer a portable language like Rust, Go, or even VM-based options such as Java or C#?”

Captain Jean-Luc PicardI’ve been having discussions with people at work about MCP. This post made me think that I haven’t been clear about my attitude toward it. I think it would be amazing if we could pose natural language queries to machines and have them do things for us. Absolutely. Who doesn’t want to just order “Tea. Earl grey. Hot.”?

My doubts are more specific to MCP itself, technically, as a protocol. This article is highly technical, but it boils down to: MCP is such a hype-y protocol right now and it’s so technically shaky that we have a responsibility to not just grab the first damned thing that shows up and make it the standard.

We did that with JavaScript and it took 2 years until it was everywhere and over 20 years until it was an actual professional tool. I’m an old man and, looking back, very often our industry is just stepping on rakes that are right there.

I just to clarify that I’m pushing back on the implementation not the idea.