I imagine a simple Bad Request could derail the convo since it's added to the context?
Fundamental things are keeping me away
- Me: "Do this"
- LLM: "Ok, let me try do this.. nope something's off. Let me try another way".
Occasionally, I do have to abandon a run but in general it seems successful.
I wonder why there is a protocol at all for end users since you are using natural language as the ACTUAL protocol.
Is MCP not for developers?
I thought of it as like a GraphQL in that it takes over the network comms layer but for LLM queries instead of payloads.
Idk
Check this out https://smithery.ai/
I see - fundamental misunderstanding on my part about what it is.
It's a chat thing - for the end user. The word "protocol" is very confusing here.
Makes sense why it didn't make sense, if that makes sense.
I invented something like this months ago to pull up a calculator, browser, etc. in the chat UI (try it chat.blacksand.cloud)
If you prompt what is "2 + 2" it will respond with "4" and render a calculator tool in the chat. It knows what tool to render based on this system.
I guess that is what they are formalizing.
I get it now, cool.
"I got an alert for X, who can help me" surfaces a bunch of helpful stuff. Not groundbreaking in the sense that I could do the leg work myself, but fantastic from automation point of view.
What scares me is the "YOLO mode" and similar in cursor. It's not practical to review each and every tool execution after the first couple of manual tool executions... Prompt injection is still a thing, whether intentional or accidental. Little Johnny Tables is still around.
It's still early days, the complicated setup with OAuth and or API tokens will be improved imo.