I Saved 90% on My Claude Code Bill With 6 Free Tools

If you use Claude Code, you know the problem.

Plans go up to $200/month and you still burn through your limits faster than you expected. The Anthropic subscription feels like renting an apartment where the rent keeps going up and you still run out of hot water.

Someone on Reddit found 6 free open source repos that actually moved the needle. Token savings of 60-90% on real projects. All free. All open source. Here's what they are.

1. ccusage — See where every token goes (15k stars)

You can't fix what you can't see.

ccusage breaks down your Claude Code usage by model, by agent, by conversation. It shows you exactly where your tokens are going — which projects, which sessions, which contexts.

Most people have no idea they're burning through tokens in one project while another barely registers. This is step one. You can't optimize what you can't measure.

Link: github.com/sourcegraph/ccusage

2. RTK — Compress terminal output before it hits the model (~60k stars)

This is where the real savings live.

RTK (R Terminal Kit, or really just "reduce the noise") compresses the output of terminal commands before they reach the AI model. It strips empty lines, groups repeated output, collapses redundancy.

The claim is 60-90% savings on command tokens alone. If you're running a lot of terminal commands through Claude Code — which you probably are — this is the single highest-impact change you can make.

Link: github.com/rtk-rnjn/RTK

3. Caveman Claude — Reply in minimal style (the gimmick that stuck)

This sounds like a joke. It isn't.

Caveman Claude makes Claude reply in a stripped-down, minimal style. Short sentences. No preamble. No "Sure, here's what I found..." Just the answer.

The developer who tried it said it felt gimmicky. He kept it on permanently.

Savings: roughly 75% per reply. If you know your project well, the short answers are just as useful. You stop paying for the prose.

Link: (search "caveman claude" on GitHub)

4. Karpathy's skills repo — Stop Claude from touching files it shouldn't

Doesn't save tokens directly. But it stops Claude from making wrong assumptions and editing files it shouldn't touch.

Which means you stop paying for the back and forth of Claude breaking something and you having to fix it. That's hidden token cost that nobody talks about.

Link: github.com/karpathy/skill-agents

5. Graphify — Local knowledge graph of your codebase (~60k stars)

Builds a local knowledge graph of your codebase so Claude checks the graph instead of re-reading your entire project every session.

Runs entirely locally. No API calls. No per-token cost. Once it's indexed your project, subsequent sessions are much cheaper because Claude isn't re-reading the same files over and over.

Link: github.com/Asabeneh/Graphify

6. Obsidian skills — Same idea, for your notes and documents

Same concept as Graphify but for your notes and documents instead of code.

If you're using Claude to reason over your notes, docs, or research, this means you build a local index once and Claude queries it instead of sending everything through the API every time.

Link: (search "obsidian skills claude" on GitHub)

How the savings compound

These aren't incremental. They compound.

RTK cuts your terminal output by 60-90%. Caveman Claude cuts your reply size by 75%. Graphify eliminates re-reading overhead. ccusage tells you which projects are burning the most so you can apply the others where they matter.

Stack them together and you're not shaving 10% off your bill. You're cutting it dramatically.

The biggest surprise from the Reddit thread: Caveman Claude. Everyone thought it was a gimmick until they tried it.

Try it.