Most people hit their Claude limit and start rationing — shorter prompts, fewer questions. That’s the wrong fix. Four hidden leaks are draining your tokens behind the scenes, and four free tools plug them in about 15 minutes.
Short answer
- Four leaks: tool output, fluffy replies, overbuilt work, and grunt work at frontier prices.
- Four free fixes: RTK, Caveman, Ponytail, and routing to free models.
- My own tests: 82.9% off tool output, 69% off reply tokens, ~half the work volume.
- Plus eight free habits — /clear, /compact, a slimmer CLAUDE.md and a memory vault.
You’re rationing the wrong thing
When people hit their limit they start typing shorter prompts, asking fewer questions, rationing tokens like water in a desert. That’s the wrong fix, because the tokens you type are a tiny slice of the problem.
The real problem is four hidden leaks draining tokens behind the scenes. Plug all four and your tokens go about five times further — same subscription, same quality, far more work.
| The leak | What it costs you |
|---|---|
| Tool output | Raw command output floods back into the conversation — giant logs Claude never needed to read |
| Claude’s replies | Three polite paragraphs where three words would do |
| Overbuilt work | You ask for one thing and get five extras “just in case” |
| Grunt work at frontier prices | Simple tasks a free model could handle, sent to your expensive plan |
Plug 1: RTK filters what comes in
RTK sits between your agent and the commands it runs. Think of a filter on a tap — the water still comes through, the junk doesn’t. It rewrites commands to be compact, filters the output, strips duplicate noise, and hands Claude only the signal.
The project claims up to 90% fewer tokens on tool output. On my own test it cut token usage by 82.9%, so their claim held up almost exactly. It adds about 14 milliseconds, which you will never notice.
Won’t filtering break things? No — it only strips mechanical noise: repeated lines, padding, the stuff Claude never needed. In my testing the quality of results stayed the same while the bill dropped by over 80%.
Plug 2: Caveman shrinks what comes out
Caveman does what it sounds like — it makes Claude talk like a caveman. Short, direct, no fluff. Ask a simple question normally and you get “Great question, I’d be happy to help, let me walk you through it”. All tokens. Yours.
In one example a normal reply used 69 tokens; the caveman version used 19. Same information, less than a third of the cost. Tested on Fable 5: 69% fewer output tokens, cutting overall usage by about 30% — and every answer was still correct.
If you like Claude being friendly, fine, for a chat. When it’s working through 50 tasks in a row you don’t need 50 rounds of pleasantries.
🔥 Want this set up without the guesswork? Four repos, about 15 minutes, then it runs automatically forever — or skip the setup entirely. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
Plug 3: Ponytail stops overbuilding
When Claude builds something it loves adding extras — layers you didn’t ask for, structure for a future that may never come. That’s tokens now, and more tokens every time Claude reads that work back later.
Ponytail makes it behave like a senior lazy developer: someone who’s done this a thousand times and knows the shortest safe path. Builds what you asked for, nothing more. About 54% less output volume, roughly 20% cheaper, 27% faster in testing.
It doesn’t cut corners where it matters — it doesn’t touch safety basics, error handling or checks. It skips the speculative “maybe we’ll need this someday” extras. And because there’s less to read back later, every future session gets cheaper too. It compounds.
Plug 4: route grunt work to free models
This is the one that changes the game. OmniRoute is a free gateway running on your own machine that routes grunt work to free models instead of your paid plan.
Think how you’d run a business. You don’t send your best person to sort the mail — but that’s exactly what most people do with Claude. Renaming things, simple summaries, tiny fixes, all sent to the most expensive model on the planet.
It gives you around 90 free models to route through. Simple task, free model. Hard task, Claude. Your paid tokens only get spent where they earn their keep.
And the clever bit: OmniRoute has RTK and Caveman built into the gateway itself, so even requests going to free models get squeezed by another 15–95%. That matters because free models have rate limits too. Setup in the OmniRoute install guide.
What stacking all four looks like
The savings don’t just add up, they multiply. RTK cut my tool output by 82.9%. Caveman cut reply tokens by 69%. Ponytail cuts work volume by about half. OmniRoute takes a whole chunk of tasks off your paid plan entirely.
Run all four and depending on what you’re doing you can end up past that 80% mark — on some workloads well past it.
Before this stack, every session had the same shape: start fast, get flowing, watch the context bloat, hit the limit, stop, wait. Plus that constant token anxiety where you’re scared to ask anything because it might cost you the rest of your session. Now it’s the same subscription, more agents at once, more done per session, and the anxiety is gone because the plumbing is fixed.
Be realistic about the numbers. These are my own measurements on my own workloads, not independent benchmarks. Tool-heavy sessions with lots of commands save the most, because that’s where RTK shines. If every task you run genuinely needs a frontier model, the routing helps less. 80% is not guaranteed on every task — measure your own before and after.
Eight extra habits that cost nothing
- Run
/clearwhen you finish a task. Old context is pure leak. Wipe it so the next task starts fresh. - Run
/compacton long single tasks. Summarises history instead of hauling the whole thing along. - Put your CLAUDE.md on a diet. Every rule in it costs tokens every single session. Claude’s own team published a guide where they cut their system by around 80% with the same quality out.
- Route by difficulty. Before each task ask whether it genuinely needs the best model on earth. Usually no.
- Plan before you build. When it plans first it wanders less, and wandering is tokens. A two-minute plan saves a pile of retries.
- Batch your asks. Five small requests means re-reading the same context five times. Send them together: one read, five jobs.
- Use a scout. Send a sub-agent to dig through the big messy stuff. It burns its own context and reports back a conclusion, so your main session pays for the summary, never the search.
- Use a memory vault. Re-explaining who you are at the start of every session is one of the sneakiest leaks going.
Old way versus new way
| Rationing | Leak-proofing |
|---|---|
| Type shorter prompts | Fix the plumbing once |
| Ask fewer questions | Ask more, they cost a fifth |
| Anxiety before every task | Run five times the tasks |
| Wait for resets | Keep working |
| Use AI less on purpose | Use AI far more |
This isn’t about using AI less. It’s the exact opposite. When each task costs a fifth of what it used to, you run five times the tasks.
And on “why not just buy a bigger plan” — you could, but you’d be paying full price to have logs padded, replies fluffed and simple tasks sent to an expensive model. You’d still hit limits, just later, having burned more to get there. Fix the leaks first, then decide whether you even need the bigger plan. Most people find they don’t.
Want the whole stack pre-built? The token-minimisation stack — all four repos preconfigured and working together with Claude, Hermes and any other agents — is inside the Agent OS in the AI Profit Boardroom, with the install file, a video walkthrough, a full token-minimisation playbook, daily updates and four coaching calls a week where you can get your setup fixed live. Start free with the free AI course and community.
FAQ
How do I reduce Claude Code token usage?
Plug the four leaks: filter tool output with RTK, shrink replies with Caveman, stop overbuilding with Ponytail, and route grunt work to free models through a gateway like OmniRoute.
How much can I actually save?
On my own tests: 82.9% on tool output, 69% on reply tokens, roughly half the work volume. Stacked, that can pass 80% on some workloads — but it depends heavily on what you run.
Does filtering output break anything?
In my testing, no. RTK strips mechanical noise like repeated lines and padding, not the signal, and result quality stayed the same.
Is it hard to set up?
Each one installs with roughly a single command, and the whole stack takes about 15 minutes. After that it runs automatically and you never touch it.
What’s the single biggest win?
Install RTK first — it’s the biggest single saving and you’ll notice it in your first session.
Why not just buy a bigger plan?
You’d be paying full price for padded logs, fluffy replies and simple tasks on an expensive model — and you’d still hit limits, just later.
What does /clear actually do?
It wipes old context so the next task starts fresh instead of dragging dead weight forward. Run it whenever you finish a task.
Will this work for other agents?
Yes. Run the compression at the gateway level and every agent you use benefits, not just Claude Code.
The bottom line
Your tokens were never the problem — the leaks were. Filter what goes in, shrink what comes out, stop the overbuilding, and send the grunt work to free models. Roughly fifteen minutes of setup, then it runs forever, and you stop rationing and start actually using the thing.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.
