Stop Wasting Tokens on AI Agents... Do This Instead!

Julian Goldie (0:00)

So great question by Bavia here, which is some basic questions. So Bavia says, you know, hello, everyone, really excited to start working on the H&OS since I don't want to run on perfection like every time, then I'm failing, I'm going to do it right this way time.

I already asked some community members how to set up. So first question here, I have chat, should I go for that or should I go for code? Great question.

It's good that you asked this today, actually. If you'd asked me a month ago, I would have said Claude, but if you already have Codex, like ChatGPT is so much better right now. So I would stick with ChatGPT. These things change very quickly. So we've all seen that, for example, when Astra 6 was released, like that literally changed everything last week. But for me personally, if I had a choice between ChatGPT and Claude today, I would use ChatGPT. Next question. I already have a little bit of experience working with VPS, because in the past I tried using this. So we have an Agent OS, right?

With an Agent OS, I wouldn't recommend using VPS for multiple reasons. Number one, I think it complicates things when you don't need to. And number two, I don't think it's that great for security, right? So those are two reasons that I don't use a VPS personally. There are members inside the AI Powerful Boarding who have set up a VPS with the Agent OS, but I wouldn't recommend that for you. In fact, if you want to connect an Agent OS to multiple devices, and you don't need a VPS, but you just want to connect the Agent OS to multiple devices, then you can use Tailscale. I wouldn't even recommend going straight to Tailscale until you set it up on one device. So make sure it's configured and working on one device, and then you can look at Tailscale later, which usually takes about 10 minutes to figure out. So if we go to the classroom, and then we go to the new daily update section, we have a full training on how to use the Agent OS with Tailscale. So yeah, don't think you need to use the VPS at all.

And then what else we got here? And then also, Baviar asked about how to reduce the amount of tokens you use, right? And that was actually what brought Baviar to the AI Profile Boarding. So we have a Token Minimization Playbook, which I'll link to for you, and that helps you reduce the amount of tokens you use massively. So we use four different free open source GitHub repos, which are Headroom, RTK, Caveman, and Ponytail. They all minimize the amount of tokens you use inside an agentic operating system. I'll link to the full training and playbook on that. But basically, these are each free repos you can use.

And you don't just have to use them with Agent OS. They work with any system. So you could plug this into Cloud Code, you could plug them into Codex, whatever you prefer. Whatever is easiest for you. So if we have a look, for example, at Headroom, it's a powerful tool that reduces the amount of tokens you use by like 20% for your coding agents, which is pretty cool. Then you have something like Ponytail, Ponytail.

He's a kind of like a lazy senior developer who will just minimize the amount of tokens that you use as well by just creating minimal code. And then you have stuff like Caveman. Caveman reduces your output tokens by about 65% by basically making it less verbose, right? So when it's replying to you, it cuts down the amount of tokens it uses and keeps it like really minimal. So these are all different playbooks that you can use to reduce the amount of tokens you use. Now, on top of that, not only this, but you can actually use an agentic operating system for free.

So you can also use an agentic operating system for free, just by using free APIs and free systems. And I'll walk you through some of the best ways you can do that. So if we have a look at the agent operating system that we have here, we can use, for example, free APIs inside Hermes, like I have a separate profile for local models, and then I can use something like LFM, 2.5, 26B.

That's a free API that is designed and was trained on Hermes agentically, so you can run that locally as an example. I ran that on a Mac Studio and it was incredibly fast, like way faster than I expected. So in terms of free options, I'll give you some ideas. If you have a DGX Spark, you could also use a local setup inside it. So if we go to the local machine, you can use free local models inside the local section here, or you could even use something like Gemma 4 Gemma 4 is a free local model that you could use inside the system. Again, you do need a decent setup like DGX Spark or RTX, but that's basically it. And then the final option is that you can use a free AI Coder. So for example, you can switch between nine router, OmniRoot, there's a brand new free API on OpenRouter as well. So those are three different free options that you can plug into the free AI Coder here.