Extra context for this project submission...
I submitted at 23:58 on the deadline day and didn't get to fully tune my readme the way I wanted, so I'm adding some additional context here:
I had to rush the demo server a bit, so the token reduction in the demo isn't as drastic as a real world scenario - but still significant. The full tool descriptions in my demo were not as verbose as most real world tool descriptions since the tools were very simple for the demo and that created less of a "wow factor" in the comparison showing the reduction.
Also, I know that I'm not the first to think of lazy loading tool descriptions but the differentiator from what I've seen is that I keep it all within the MCP spec and my method (putting the descriptions right in MCP resources) is simpler than most of the others I've seen proposed. From my research, most lazy loading proposals are more tool specific and not general MCP enhancements.
I also think most overlook the importance of a fallback / enforcement mechanism for ensuring that when the LLM does use tools, it fetches their full descriptions first. Using the authorization methods already laid out in the MCP spec allows us to do this in a more deterministic way versus relying solely on the system prompt.
Anyway, let me know if you have any questions.
Hi there! Just a quick reminder that our deadline was November 30, UTC end of day. We are not accepting submissions after that. Thanks for understanding!