Raw LLM Responses

I like the phrase "thinning layer" - clever way to say it. I will say though that Anthropic is edging towards a traditional UI as these "MCP Apps" come into play. The latest Claude Desktop update started rendering these UI elements (though honestly I'd love to be able to toggle them off at times; not quite ready for primetime). I don't expect us to jump back to a chat prompt to do everything; nobody is pining for the DOS C: prompt as the only means to interact. But there's a great deal of advantage of having a flattened data layer, not the least of which is having no boundaries between your data...which does necessitate that you have control of it (either via trusted connectors, or simply possessing it). Agreed re: the Karpathy Wiki. What I realized when building the plugins for FlashQuery is that's where behavior was defined. You want a multi-connected knowledge base that self-connects over time? The plugins can manage that behavior, but you still need the data plumbing under the hood. The app is a plugin. I just have a sense that with a database, versioned markdown, and skills...you get get a lot done. Re: permissioning: Good question. That's why I didn't make this \*enterprise\* ready. But I do think we'll need a capability oriented model for handling permissions, including user-delegated actions that agents can run with, and re-delegate...all cryptographically signed.