4 FREE Ways to Save 80% on Claude Claude Tokens!

**Julian Goldie** (0:00)
Today, I'm going to show you four free GitHub repos that cut Claude Code tokens by 80%.
So this is a full token minimization playbook that cuts your Claude Code tokens by 80%, and all of it is free. Now, you can use this with any AI agent, it doesn't have to be Claude Code. And these are four free tools, each one reducing the amount of tokens you use, so you can get more out of your AI agents. That means your AI agents get faster, your limits stop running out, and everything keeps working with the setup you already have. And I've actually tested every single one. In fact, one of them actually cut a single command's output by 92%, we'll come on to that later. So that by the end of this, you'll have the whole stack installed, and your tokens will go five times faster, five times further. So the problem with Claude right now is like, most of the time we're using tokens without being smart, right? So Claude gets flooded, you get like giant test logs, it didn't need to read, you've got a polite three paragraph answers, overbuilt code nobody asked for, right? And also like full jobs being done, which a free model could actually do for you, right? So it takes four different tools, and all free open source repos that you can use to get more out of this stuff, right? And so if you're wondering, okay, what are the problems with this? Well, leak number one is tool output. And so one of the biggest problems is the clored replies way too politely with loads of filler words, which reduces, which increases your output tokens, which uses up your token plan. Also another problem is overbuilt code, right? So sometimes it overbuilds the code that it gives you, and it's full of fluff that doesn't need to. Additionally, you can end up paying full price for grunt work, and it renames variables and write stuff that it doesn't need. And also tool alpha, right? It sometimes floods with context. So these are all things, all different leaks that sneak in to clored and cost you more tokens. And so you might say, well, is this actually worth the whole setup time? The whole stack can install in about 15 minutes with one command each. And after that, it's pretty much automatic. You don't need to do it again. We've actually got it running inside our agent operating system. So, for example, with clored, we can use these token minimization playbooks to reduce the amount of tokens we use each time.
So what is the token minimization playbook? Well, there's four parts to it. So we have RTK, caveman, ponytail and Omniroute. Now RTK is basically a powerful way to reduce the tool output, and it can actually reduce 89.9%.
Based on my own tests, it reduced the token usage by 82.9%.
Basically what this means is like when it is running shell commands, usually it needs a fraction of what it actually uses. And the problem is like if it's using a git diff, which you don't need to know about, but basically it's a way of doing shell commands. If it's messy, it can use up thousands of tokens. So RTK sits between the agent and the shell command, which rewrites the command to compact it, filter the output, de-duplicate the noise, and hand-clawed only the signal. It adds about 40 milliseconds, so it doesn't slow it down. You don't see it working. It just reduces the amount of tokens you use. Now we actually tested it because on RTK, they claim 90 percent fewer tokens. And actually on my tests, it reduced 82.9 percent. So it was very, very accurate and very close. It's a free open source project. This is RTK, as you can see right here. Basically, it allows you to reduce the amount of tokens we use. Now we've already got it built into the agent OS system with all of our agents running together. So we have Claude ready to go to reduce the amount of tokens we use.
You might also say, well, won't filter in the output break the agent. So it only filters mechanical noise.
That's all it's filtering out there. Then we have Caveman. Caveman is the second part of the system, the second free repo. And Caveman basically reduces the output tokens that it speaks with, right? So it can actually reduce tokens by 65%.
So you can see an example right here. And if you look at the before and after, a normal agent would use about 69 tokens on an output. Whereas for example, with Caveman, we reduce that to 19 tokens, which means that we use way less tokens per output. And instead of these fluffy filler responses that are super polite, it just gets straight to the point, which is what we want. And we use less tokens that way as well. And from my tests on Fable 5, it used 69% fewer output tokens, which reduce the overall tokens I was using by 30%.