Speaker unverified: belief

Transcript context

…That's where I think the concepts of resources and tools get to blend a little bit, right? Because now you're saying you want some degree of user control, right? Or application control. And other times you want the model to control it, right? So now we're choosing just subsets of tools. I don't know. Yeah, I think it's a fair point or a fair concern. I guess the way I think about this is still like at the end of the day, and this is a core MCP design principle is like, ultimately, the concept of a tool is not a tool. It's a client application, and by extension, the user. Ultimately, they should be in full control of absolutely everything that's happening via MCP. When we say that tools are model controlled, what we really mean is like, tools should only be invoked by the model. Like there really shouldn't be an application interaction or a user interaction where it's like, okay, as a user, I now want you to use this tool. I mean, occasionally you might do that for prompting reasons, but like, I think that shouldn't be like a UI affordance. But I think the client application or the user deciding to like filter out the user, it's not a tool. I think the client application or the user deciding to like filter out things that MCP servers are offering, totally reasonable, or even like transform them. Like you could imagine a client application that takes tool descriptions from an MCP server and like enriches them, makes them better. We really want the client applications to have full control in the MCP paradigm. That in addition, though, like I think there, one thing that's very, very early in my thinking is there might be a addition to the protocol where you want to give the server author the ability to like logically group certain primitives together, potentially. Yeah. To inform that, because they might know some of these logical groupings better, and that could like encompasses prompts, resources, and tools at the same time. I mean, personally, we can have a design discussion on there. I mean, personally, my take would be that those should be separate MCP servers, and then the user should be able to compose them together. But we can figure it out. Is there going to be like a MCP standard library, so to speak, of like, hey, these are like the canonical servers, do not build this. We're just going to take care of those. And those can be maybe the building blocks that people can compose. Or do you expect people to just rebuild their own MCP servers for like a lot of things?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.