Post

okgodoit (Roger)
@okgodoit.com

I’ve read through the spec and the main thing that’s confusing to me is this seems to largely assume that clients will run locally on your own computer? There’s mentions of SSE but I don’t see any implementations or any way to point Claude at a url like example.org/.wellknown/mcp for example.

okgodoit (Roger)
@okgodoit.com

From a UX point of view it seems like a big ask to have end users install programs or scripts on their computer that Claude desktop then talks to. Seems like pasting in a URL or having some sort of directory of URLs would be a better way to do this. What am I missing?

Simon Willison
@simonwillison.net

Completely agree - I think this will all become a whole lot more interesting and usable when they add it to the Claude web app and mobile apps using their HTTP+SSE based transport option

Tony
@anthonypowell.me

So if my understanding is correct: Tools and Resources need to be fetched and advertised to the LLM, through a transformation step to turn their responses from the MCP client into tool definitions. Then tool calls need to be routed back to the MCP client, and those results, need to go back to llm

Natik Gadzhi
@natikgadzhi.bsky.social

I’m very much looking to spend the holidays on trying to make it work with Airbyte api data connector format, and see if I can make models benefit from data on tap from arbitrary APIs. 👀 Haven’t felt that excited for a good quarter.

mr. TIM
@timkellogg.me

imo the spec seems too big to be a good standard. like, it’s launching 1.0 with custom transports? idk, smells like design by committee

Unclecode (Hossein)
@unclecode.bsky.social

This definitely outshines OpenAI GPTs! You can wrap any tools into a simple server app, pass it to Claude, and that’s it. Claude can use them all! It’s far more flexible and, with good design, allows for scalability. To me, this is one of the best moves in increasing LLM utility in recent years.

Simon Hürlimann
@huerlisi.bsky.social

Did go through the docs and I see some potential. Might allow plugging in services into central Apps like Claude Desktop (or open-interpreter etc.) and IDEs. But quite a bit more complex than tool calls and dedicated context building in "normal" apps that provide Chat/Agent based workflows.