If you've been searching for how to reduce Claude Code token usage, here's the punchline up front: I cut mine past the 80% mark on some workloads using four free tools, and I never typed a single shorter prompt to do it. The tokens you type were never the real problem. The leaks were — and I'll show you where they hide and how I plugged all four in about 15 minutes.
When most people hit their limits, they do the obvious thing: shorter prompts, rationed sessions, less AI on purpose. It feels sensible. It's the wrong fix, because your typed prompts are a tiny slice of what actually burns through your plan.
How To Reduce Claude Code Token Usage: The Four Leaks
Nearly all the waste comes from four hidden leaks:
- Tool output. Every command your agent runs dumps its raw output into the conversation — giant logs Claude never needed to see, all billed to you.
- 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" — and every extra costs you again on every future read-back.
- The biggest: the wrong model. Paying frontier-model prices for grunt work a free model could do in its sleep.
Plug those and rationing becomes irrelevant. Below is the stack I use — one free tool per leak, each installing with roughly one command from its own repo, about 15 minutes for the lot, then automatic forever. It's the exact setup wired into my Agent OS, the free coding engine I run everything through.
RTK: Plugging The Tool Output Leak
RTK sits between the agent and the commands it runs, like a filter on a tap. It rewrites commands to be compact, filters their output, strips duplicate noise, and hands Claude only the signal. Ask Claude to check what changed in a folder and you'd normally flood the context with a huge log; with RTK, Claude sees the short clean version instead.
The project claims up to 90% fewer tokens on tool output. I don't repeat other people's marketing numbers, so I measured it myself — the same test-it-yourself habit my Goldie Bench runs are built on. My result: an 82.9% cut, adding about 14 milliseconds per command. Unnoticeable. And it only filters mechanical noise — repeated lines, padding — so the signal gets through. In my testing, quality stayed the same.
Caveman: Plugging The Reply Leak
Caveman makes Claude talk like a caveman: short, direct, no fluff. A normal reply opens with "Great question! I'd be happy to help you with that…" — all tokens, no work. One example reply came in at 69 tokens normally and 19 with Caveman. Same information.
My test on Fable 5: 69% fewer output tokens, which cut my overall usage by about 30%, and every answer was still correct. For a casual chat, keep Claude friendly. For 50 tasks in a row, you need the work, not 50 rounds of pleasantries.
Ponytail: Plugging The Overbuilding Leak
Ponytail makes Claude act like a senior lazy developer — someone who's done this a thousand times and knows the shortest safe path. It builds what you asked for and nothing more. My numbers: about 54% less output volume, roughly 20% cheaper, 27% faster in testing.
It does not skip safety basics, error handling or checks — only the speculative "maybe someday" extras. And because there's less to read back later, every future session gets cheaper too. It compounds.
Omniroot: Plugging The Wrong-Model Leak
Omniroot is the big one: a free gateway on your own machine that routes grunt work to free models instead of your paid Claude plan. You don't send your best person to sort the mail. Renaming, simple summaries, tiny fixes — there are 90 free models to route through. The hard tasks stay with Claude.
The clever bit: Omniroot has RTK and Caveman compression built into the gateway itself, so even requests headed to free models get squeezed another 15–95% — which matters, because free models have rate limits too. I've published a full Omniroot install walkthrough and a guide to running it with a Hermes agent if you want the deep dive.
If you want this whole stack running without touching a config file, the AI Profit Boardroom ships the whole leak-proof stack pre-wired inside the Agent OS. → Get the pre-wired stack
The Stack At A Glance
| Leak | Tool | What it does | My measured result |
|---|---|---|---|
| Tool output flooding the context | RTK | Filters command output, strips duplicate noise, keeps the signal | 82.9% fewer tool-output tokens, about 14ms added |
| Padded replies | Caveman | Short, direct answers carrying the same information | 69% fewer output tokens, about 30% off overall usage |
| Overbuilt work | Ponytail | Builds exactly what you asked, no speculative extras | About 54% less output volume, 20% cheaper, 27% faster |
| Frontier prices for grunt work | Omniroot | Routes easy tasks to free models, keeps hard ones on Claude | Whole task categories at 100% off, plus 15–95% built-in compression |
📺 Watch: How to Reduce Claude Code AI Tokens by 80% (FREE)
What Stacking Them Feels Like
Here's the part most people miss about how to reduce Claude Code token usage: the savings multiply, they don't add. Take 82.9% off tool output, 69% off replies, roughly half the work volume off the top, then send entire task categories to free models at 100% off. On some workloads that lands past the 80% mark; on tool-heavy sessions, well past it.
The before-and-after is stark. Before: sessions started fast, bloated, hit the limit and stopped — with token anxiety humming in the back of my head all day. After: same subscription, more agents running at once, more done per session, anxiety gone.
Eight Habits That Cut Claude Code Tokens Even Further
The tools fix the plumbing. These habits fix you:
- Run /clear when you finish a task. Old context is pure leak — you're paying to carry a conversation you no longer need.
- Run /compact on long single tasks. It keeps the gist and drops the bulk.
- Put your CLAUDE.md on a diet. That file rides along with every session, and every rule costs tokens every time. Most people can delete half of it and nothing breaks — Claude's own team put out a guide where they cut their system prompt by around 80% with the same quality.
- Route by difficulty. Ask yourself: does this actually need the best model on Earth? Usually no. A free API covers more than you'd think, and routers like 9Router make the handoff automatic.
- Plan before you build. A two-minute plan saves a pile of retries. Wandering is tokens.
- Batch your asks. Five requests sent together means one read of the context and five jobs done, instead of five separate re-reads.
- Use a scout. A sub-agent digs through the big messy stuff in its own context and reports back just the conclusion. Your main session pays for the summary, never the search.
- Use a memory vault. An Obsidian memory hooked to your agents means never re-explaining who you are each session — re-explaining is one of the sneakiest leaks going. My guide to the best memory setup covers it.
📺 Watch: Jcode DESTROYS Claude Code?
The Honest Caveats
I want this to be straight with you, so here's where it bends:
- Routing helps less if every task genuinely needs a frontier model. But be honest about how much of what you send actually does. For me, it's a fraction.
- 80% isn't guaranteed on every task. Tool-heavy sessions save the most; lighter sessions save less.
- My numbers are what I measured — 82.9% on tool output, 69% on replies, roughly half the work volume — not projections. Your workload will land somewhere different.
- "Just buy a bigger plan" doesn't fix it. You'd pay full price for padded logs and fluffed replies, and still hit the limit later. Fix the leaks first, then decide whether you even need the bigger plan.
📺 Watch: Jack Dorsey's Buzz + Claude Code is INSANE!
The Old Way vs The New Way
The old way treats tokens like water in the desert. You ration, you hover over every prompt, you deliberately use less AI — and a low-level anxiety hums behind every session.
The new way fixes the plumbing once — 15 minutes, four free repos — and then you do more, not less. When each task costs a fifth of what it did, you don't pocket the difference. You run five times the tasks. The limit was never really the problem. The leaks were.
The Install Order
If you're starting from zero on how to reduce Claude Code token usage, order matters. Install RTK first — it's the biggest single win and you'll see it in your very first session. Then Caveman, then Ponytail, then Omniroot with routing switched on. Each one is about one command from its repo. Then layer the eight habits on top over your next few sessions.
FAQ
Does filtering the output break the agent?
No — this was my biggest worry too. RTK strips mechanical noise only: repeated lines and padding. The signal always gets through, and in my testing Claude's work stayed identical while tool-output tokens dropped 82.9%.
Is this actually free?
Yes. All four are free open-source tools, and the models Omniroot routes to cost nothing — the catch is that free models carry rate limits, which is exactly why the built-in compression matters. Pair the stack with my guide to running Claude Code free and your only real cost is the 15-minute setup.
Will the quality of the work drop?
Not in my testing. Caveman delivers the same information in fewer words, and every answer was still correct. Ponytail skips speculative extras, never error handling or checks. RTK removes noise Claude didn't need anyway.
Do I still need a bigger plan?
Fix the leaks first, then decide. Upgrading before you fix them means paying full price for padded logs and fluffed replies — and hitting the limit again later anyway. Most people who plug the four leaks find the plan they already have suddenly feels enormous.
What saves the most?
On tool-heavy sessions, RTK — that's where my 82.9% came from. Across everything, routing — because sending a task to a free model is a 100% saving on that task. Stack all four and let the multiplication do the heavy lifting.
The Verdict
That's how to reduce Claude Code token usage in practice: stop rationing and fix the four leaks. Fifteen minutes, four free repos, and the token anxiety that used to shape my whole workday is gone — same subscription, far more output. The limit was never the villain. The plumbing was.
If you want your Claude Code plan to stop being the bottleneck in your business, check out the AI Profit Boardroom — inside you get the token minimisation stack pre-built into the Agent OS with the install videos and the full playbook, four coaching calls a week where your Claude Code setup gets fixed live, daily updates, and 3,700+ business owners running these agents. → Plug the leaks and run more agents
