7 Free Repos That Cut Your Claude Code Token Bill
TL;DR
Every command you run in Claude Code burns tokens you never see - file dumps, long diffs, bloated answers. These free repos make each prompt cheaper: measure what you spend, compress what goes in and comes back, and stop paying to re-read the same files. Install the first one in the next ten minutes.
Why your bill creeps up
Every command you run eats tokens. A single file dump or a long git diff can push thousands of tokens into the chat and you never see it happen - you just hit a rate limit faster, or the bill is bigger than you expected. The fix is not to use Claude Code less; it is to make each prompt cheaper. Six of these repos cut tokens directly, the seventh teaches the habits that cut them on their own, and every one is free.
See what you actually spend first
Reads the usage files Claude Code already saves on your machine and shows what you spent by day, week and session in real dollars - no login, no setup. Run `npx ccusage daily` and you will spot your biggest token leak in under a minute. You can't cut what you can't see, so start here.
Cut tokens on both sides of the conversation
The two highest-leverage installs work as a pair: one shrinks what you send Claude, the other shrinks what Claude sends back. Add them together for the biggest, fastest cut.
Sits between your terminal and Claude and compresses bulky command output - file listings, file reads, test runs - before Claude ever sees it. Install with `brew install rtk && rtk init -g`. Typical 30-minute sessions shed around 80 percent of their tokens; one real run dropped from 118,000 tokens to 23,900.
Where rtk shrinks what goes in, caveman shrinks what comes back. It is a skill that tells Claude to answer in tight, dense language with no filler while keeping every technical detail, and it trims your memory files so the saving compounds on every session after. Around 65 percent fewer output tokens with no loss in accuracy.
Stop re-reading the same files
Every time you ask about your codebase, Claude re-reads the files and you pay for it again. graphify reads your folder once, builds a map of how everything connects, then answers from the map. Install with `pip install graphifyy && graphify install`, then run `/graphify .` in Claude Code. It only rebuilds files that changed, so it stays cheap.
Clean the web before Claude reads it
When Claude reads a web page it pulls in the ads, menus and junk too, and you pay tokens for all of it. The defuddle skill in this pack strips a page to clean markdown first, so Claude only reads the part that matters. Add it with `npx skills add https://github.com/kepano/obsidian-skills`. Built by the head of Obsidian and built on the open Agent Skills spec, so it works across agents.
The habits that cut tokens on their own
A running list of Claude Code tips, several of which are pure cost wins. The biggest is patching: turning it on trims the system prompt from roughly 19k tokens to 9k - about 10,000 tokens saved every single session before you type a word.
- Turn on patching. It cuts the system prompt from about 19k tokens to 9k, saving roughly 10,000 tokens every session before you type anything.
- Run /compact when a chat gets bloated. It summarises the thread and frees the context window instead of letting it fill.
- Start fresh more often. A clean conversation performs better and costs less than one long thread you keep dragging along.
- Keep your CLAUDE.md short and built on simplicity rules - think first, stay surgical, make small changes. It loads with every chat, and it stops Claude overbuilding answers you only have to redo.
- Break big jobs into small ones. Smaller tasks use less context and land right more often, so you redo less.
- Watch /usage so a rate limit never surprises you mid-task.
How to stack them
You do not need all seven on day one. Add them in this order and watch the number in ccusage fall after each step.
- 1Install ccusage and look at what you actually spend - you can't cut what you can't measure.
- 2Add rtk and caveman together. That is the biggest, fastest cut, on both sides of the conversation.
- 3Shorten your CLAUDE.md around simplicity rules so Claude stops overbuilding.
- 4Add graphify once you are working in a real codebase, and the defuddle skill once you are reading a lot of web pages.
- 5Turn on patching, then run ccusage again and watch the number drop.
Common questions
Do I have to install all seven repos?
No. Work in order: install ccusage first to see what you actually spend, then add rtk and caveman for the biggest cut, and bring in graphify and the defuddle skill only when you are working in a real codebase or reading a lot of web pages.
How do I install a GitHub repo into Claude Code?
The fastest way is to copy the GitHub link, paste it into Claude Code, and say 'install this for me' - Claude reads the repo and sets it up. Each repo above also lists a manual command if you prefer to do it by hand.
What is the difference between rtk and caveman?
rtk shrinks what goes in - it compresses bulky command output like file listings and test runs before Claude sees them. caveman shrinks what comes back - it makes Claude answer in tight, dense language and trims your memory files. Run them together to cut both sides.
Will compressing tokens hurt answer quality?
No - these tools remove noise, not signal. rtk drops terminal clutter, caveman keeps every technical detail while cutting filler, and defuddle strips ads and menus from web pages. You pay for less junk, not for less answer.
What is the single fastest cost win with no install?
Turning on patching. It trims the system prompt from roughly 19k tokens to 9k, saving about 10,000 tokens every session before you type a word - no repo required, just a setting.
Are these repos actually free?
Yes. Every one is open source on GitHub at no cost. The only thing they spend is a few minutes to install - and they pay that back the first time you check ccusage.
Want the rest of the cost-cutting stack?
Get the other 3 in the cost-cutting stack - free, with 5,000+ builders.
Join the Club