In this panel discussion, the speakers address the high token consumption and associated costs when using AI models for large context analysis. Tim explains their solution: a custom, lightweight Design System MCP delivered as a local npm package, which provides the necessary agent context without requiring remote API calls, thereby reducing token usage, latency, and authentication overhead. The conversation also highlights a hybrid approach that combines deterministic traditional machine learning algorithms with AI for B2B workflows, leveraging the strengths of each. Additionally, a brief mention is made of Headroom, a tool for compressing token usage across AI applications. Overall, the discussion emphasizes practical strategies—local context packages, deterministic‑AI hybrids, and token compression—to lower both time and monetary costs while maintaining or improving result quality.
Chapters
During AI implementation, we've encountered a problem. The more information we ask the AI to analyze, the more credits it consumes. In the end, the outcome looks the way we want, but it took a lot of credits. Have you tried to optimize the cost? Tim, you can, I know the answer to this, but you could talk a little bit more about that kind of journey with the MCP and—
Yeah, it's a good question. I mean, everybody's facing that now or will be shortly as they, as they realize that Oh, there's, there's a cost associated with this. And so when we've designed a custom MCP in the past, and when, when we will in the future too, it's something that runs locally. With this in mind, the context for the design system that an agent needs is not, is not so big that it needs to be hosted somewhere, or it's counterproductive to have your local machine download that MCP. It's relatively small. And so we've developed it as a— our custom MCP is a simple npm package that can be installed that just again runs locally. It's not something that is hosted or just leads to more consumption of tokens. And if the agent has that context at the ready and your machine doesn't need to— your computer doesn't need to do a whole bunch of processing to figure out what the answer is given the context in that MCP. If it's just looking up basically like the equivalent of looking up in a dictionary, here's the answer to the question, well, it's, it's faster time-wise. It's less of a spend because the agent's not churning through tokens. It's— you're not dealing with authentication challenges because it has to go out to a hosted solution. There's so many advantages to that, and so that's what we'd recommend is make it local., and you'll get the same quality result, maybe even a better result while keeping costs as low as it could be, both in time and dollars.
I think the other thing too is it, it goes beyond design systems, right? Even just using deterministic traditional ML with algorithms and then using AI where you need to. I mean, I think that's what we've seen as the biggest combination, especially when you're dealing with B2B software and things that are more deterministic. You know, if you think about workflows and other things, those two together. Together.
I saw there's a lot of, there's a lot of buzz around Headroom as well, which is a way to compress token usage as well, just for general, general AI usage. So something to look, look into if you haven't yet.
