Headroom: Never hit Claude’s Usage Limit Again!

**Julian Goldie** (0:00)
Today, I'm going to show you a new free powerful way to reduce the amount of tokens you use, whether you're using Claude, for example, or any sort of model like that. Obviously, it requires more tokens, and the more powerful these models become, usually the more tokens they require. So this is a new free open source project, it's called Headroom, and you can see it's already trending on GitHub. And what we've got here is basically a way to use 60 to 95% less tokens when we are running an AI agent. And you can plug this into any of your agents, they actually show proof right here. So you can see the amount saved in terms of tokens, 90 saved, 73% on a triage issue, 47% saved on a code base exploration. So whether you're using this for coding or a genitas or anything like that, you can save tokens using this process. And you can also plug it into Claude Code, into Codex, into Cursor, into OpenClaude even, and probably even as Hermes agent as well. So this is great if you run agents daily and you want to save tokens without changing your code, if you work across multiple agents and one shared memory is also useful. And also if you need reversible compression. I'm going to walk you through exactly how to use it, how to get it set up and also how to install it. So the way that I would look at this is like before your AI agent answers a single question, it quietly reads like a mountain of text. It could be your files, your tools, your search results, your skills, your logs, your error messages, the whole conversation so far, particularly if it's a long running conversation like with an AI agent. And so the problem is that every single word, every single request uses up tokens. The more powerful these agents get and the more they can do, the more tokens they use up. So agents, AI agents in general, we use up more tokens than just using a single chat, like one-to-one with, for example, chat GBT. And so if you get slow to crawl and it's costing you tokens, and particularly if you're using an API as well, then also what also happens is that it can forget what you told it five minutes ago as well. And so this free tool, Headroom, crushes that mountain of text and all that stuff that has to read by 60 to 95%.
So you get the same answers, but you don't throw anything away. This is how to use it. Now, most AI agents that people have problems with, it's not because they aren't very good. It's usually because it's just too much text and too much going on and that slows it down. We saw that with OpenClaude as well. The more updates they brought in and the more stuff they added to it, the slower the updates got and the less useful the tool became because it was getting slow, forgetful, and also a little bit buggy with all that stuff going on. And so the problem is every time your agent answers something, it first has to read everything that you've handed it. And that's usually huge. Every word it reads is a token. You pay per token. So more reading equals more tokens used and more costs and more resources. And so the other problem is as well, that your AI only has so much room to hold text at once, the context window. And so if you have a small desk and you pile too much on it and the early pages slide off the edge, your agent forgets the start of the job halfway through. And so quite often it's not even the intelligence that costs a lot of tokens, it's actually the stuff that you don't need inside your AI agent. And Headroom is like a zip file for everything your AI reads. Now if you want to install it, what you can actually do is you can plug this in to your AI agent so we can copy the information from GitHub here and say install this to use less tokens. And we'll just plug that into Hermes as you can see. And now you can see Hermes is setting up. So Headroom basically sits quietly between your AI and all the text that you have before anything reaches a model. It squashes it down and compresses it, keeping the meaning but dropping the bloat. So you get the same answer, but you get a fraction of the reading. And three things make it special. So number one, it can run in your machine. Number two, it covers everything. So files, search results, logs, conversation, history. Number three, it's actually free, right? It's a free open source project on GitHub available right now. And so the way that I look at this, something I call the Goldie Token Crusher Framework. Most people will never touch this because it sounds technical. That's a mistake because it's just a simple three step process, right? Number one is you squash everything the agent reads down to a fraction, which means 60 to 95 percent smaller without losing the meaning. So you get the same job, but less tokens required. Move number two is you keep it. So originals are never deleted. You've got one shared memory. We actually use Obsidian for that in our memory.