Make your AI coding plan go further. From the first request to the last token.
TokenMaestro is being built to conduct the work between you, your AI agent and your computer: a local model will structure your request, ask what’s missing before the agent starts, hand heavy work to tools installed on your machine, and give each task the right model and effort. Every saving will be measured on your machine, never estimated.
First release, for Claude Code: the free X-ray and, in Pro, the session guard. Next: the full orchestra, and Codex, Gemini and other agents.
Free X-ray · No account · Nothing leaves your computer
The orchestra: fewer tokens from the start
Most token tools compress what the agent has already spent. TokenMaestro works before the spend: on the request, the understanding, the tools and the model.
A well-built request
A small open model runs on your computer and turns what you typed into a clear, structured request before it reaches the agent. It costs no tokens.
Questions before the work
When a request is ambiguous, TokenMaestro asks you first, so the agent doesn’t spend tokens guessing and redoing.
Local tools for the heavy lifting
Reading text in images, converting documents, cutting video: tools installed on your machine do it, set up by TokenMaestro with your permission. The agent only gets the result.
The right model and effort
Simple tasks go to a lighter model at low effort; the heavy model and high effort only when the task needs them.
Lean sessions
Warnings inside Claude Code, each with its cost at API prices: when the context has grown large, what every further step pays to re-read it; when a break has let the cache expire, what writing it again costs over a warm turn. On resume, the warning comes before your first message; after a break mid-session, it comes as your next message goes through.
Every token accounted for
The X-ray shows where every token went, counted from the logs, never estimated. Next: each fix measured on your own work, on and off; if it doesn’t pay off, it is switched off.
Your agent, your choice: Claude Code first; Codex, Gemini and others next. TokenMaestro conducts; you choose the instruments.
Why this saves tokens
In an AI agent, every step re-reads everything that came before. A token left in at the start is paid again on every step after it, and a token cut at the source is saved on all of them. That’s why TokenMaestro works before the spend.
A clear request, fewer loops
An ambiguous request turns into attempt, fix and another attempt, and every loop re-reads the whole session. Structuring the request first cuts loops.
A local question costs nothing
Clearing up a doubt on your computer spends no cloud tokens. The agent finding out it misunderstood costs a whole round, plus the fix.
Heavy files stay out of the context
A 50-page PDF read by the agent enters the context and is re-read on every step. A local tool extracts only the part that matters.
A model sized to the task
A light model costs a fraction of the strongest one, and high effort generates a lot of reasoning text. The strong model comes in only when the task needs it.
A session that doesn’t balloon
Once the context passes a few hundred thousand tokens, every answer gets expensive. Compacting at the right time, and not letting the cache expire during a break, avoids rewriting everything at full price.
Measured, not promised
Every adjustment will be measured on your own work, on and off. If it doesn’t save, it will be switched off. Until then, no saving is claimed.
How it runs, start to finish Coming
- You install it once. TokenMaestro finds the agents you already use and the tools on your machine.
- You make the request through TokenMaestro: in the terminal, or from inside the agent itself.
- It prepares the request on your computer (structures it, asks what’s missing, runs local tools) and hands it, ready, to the agent you chose.
- During the session it watches context, cache and limits, and warns you before they turn into waste.
- At the end, the X-ray measures what changed.
Today: the X-ray reads Claude Code’s logs whether you sign in with a subscription (Claude Pro or Max) or an API key, and shows the value at API prices either way. Coming: the orchestra will drive the official apps of the agent you choose (Claude Code first; Codex and Gemini next), the way you already sign in to them.
Start with the X-ray: what your plan was worth, and where it went
One command reads months of logs in seconds and writes a single dashboard file: value at API prices by day, model, effort level and project; days you hit the limit; and a cross-check against Claude Code’s own counters.
What it finds: the leaks, with a label on every number
“Measured” means the logs show it. “Ceiling” means the most a fix could recover — not a saving. In the founder’s logs:
Huge contexts
of the value at API prices went to turns with more than 400k tokens of context. That is a share of the spend, not an amount that can be recovered. Each step re-reads everything before it.
Cache rewritten after breaks
The cache expired 491 times during pauses, and the part of the context that was no longer in the cache was written again at full price.
Tokens before the first answer
Instructions, tools and memory every session starts with — re-read on every turn after that. Counted up to 100k tokens per session.
The weight map
What your agent read the most, by project and file. Area is the volume that entered the context; color shows how much of it was re-read with no change to the file.
Next: the same map drawn over your code’s structure, with the path your agent took in each session.

Why you can trust the numbers
Most token tools report savings they estimate themselves. Independent measurements of popular ones found no savings, and sometimes higher cost. We started from the other end.
The log traps, handled
Claude Code writes several records per message, the early ones with partial counts; resumed sessions replay messages; subagents write separate files. Each case has a test.
Checked against Claude Code
In 16 comparisons, our cache-read count came out 4.6% below Claude Code’s own counters, never above. The numbers are a floor.
No savings claims until measured
No savings percentage goes on this page until a paired measurement is published with the raw data and a way to reproduce it.
Download
The X-ray is a small command-line tool. Paste one line in your terminal to install it, or download the file.
irm https://thetokenmaestro.com/install.ps1 | iexWindows SmartScreen may show a warning for the downloaded file, because the program is not signed yet. The install line above is the recommended way.
curl -fsSL https://thetokenmaestro.com/install.sh | shtokenmaestro xrayWorking in Claude Code? You can ask it to run the install line for you.
To update, run the install line again. To uninstall: tokenmaestro guard uninstall (if the session guard is on), then tokenmaestro deactivate (if you have a Pro license), and only then delete the program (Windows: and its PATH entry).
TokenMaestro runs on your computer: Windows, macOS or Linux. Open this page there to download it.
Pricing
X-ray
- Value and usage dashboard
- Leaks, with evidence labels
- Weight map
- Cross-check with Claude Code’s counters
- Missing-tools check
Pro
Founding members: $19 a year, locked in for as long as you stay subscribed. First 1,000 people.
- Session guard: a warning inside Claude Code when the context passes 200k tokens, with what re-reading it costs on every further step, at API prices
- A warning when a break has let the cache expire, with the extra cost of writing the context again, over a warm turn
- Status line for the terminal: context, re-read cost of the next step, cache and plan limits
The status line works in the Claude Code terminal; tokenmaestro guard install shows how to switch it on.
- Structured requests from a local model
- Questions before the agent starts
- Local tools installed and tuned
- The right model and effort for each task
- Savings measured on your own work
- Codex, Gemini and other agents as they arrive
The license key arrives by email. Activate it with tokenmaestro activate YOUR-KEY. It works on up to 3 computers. The session guard needs Claude Code 2.1.251 or later.
Two years of Pro cost less than a single month of the step from one Claude plan to the next.
FAQ
What does TokenMaestro read?
The session logs Claude Code already writes on your machine (~/.claude/projects): the token counts the API returned for each message, and which tools ran and how much they returned. Nothing is sent anywhere.
Is the dollar value what I paid?
No. It’s what your usage would have cost at API prices. On a flat-rate plan it shows how much your plan delivered, not a bill.
Claude Code already has /usage. Why this?
/usage tells you how much of your limit is gone. TokenMaestro tells you where it went, what it was worth, and what’s leaking.
How much will I save?
We don’t promise a number before measuring. The X-ray shows how much is leaking in your case. When fixes ship, each one is measured on your own work.
Do I need an account?
No. The X-ray needs no account and no email. Pro uses a license key, sent by email when you buy.
Does it work with Codex, Cursor or Gemini?
Not yet. Today it reads Claude Code only; Codex and Gemini come next, and Cursor is not supported for now. You choose the agent; TokenMaestro conducts the rest.
Windows, macOS, Linux?
Yes. The X-ray runs on all three.
Do I need to install models or tools?
No. When a task calls for one, TokenMaestro shows what’s missing, explains why and installs it with your permission. The local model is small and open; if your computer can’t run it well, TokenMaestro works without it. (Pro, coming.)
Does it work with a subscription, or only with an API key?
Both. The X-ray reads the logs Claude Code writes whichever way you signed in, a subscription (Claude Pro or Max) or an API key, and shows the value at API prices in both cases. It cannot show the share of your plan’s limit each thing used, because those limits are not published; it does mark the days the limit was hit. The orchestra, when it ships, will drive the official apps the way you already sign in to them and never use your login outside them.
Is there a refund?
Yes. You can ask for a full refund within 7 days of the purchase if you live in Brazil, within 14 days if you live in the EU, EEA or UK, and anywhere else within 7 days of your first purchase of a plan. Write to hello@thetokenmaestro.com. Creem, our merchant of record, processes the refund, and the license key is then disabled. The X-ray is free, so you can see what TokenMaestro finds on your own machine before you buy.
How do I uninstall it?
First run tokenmaestro uninstall: it removes the session guard’s hooks from Claude Code and frees this computer’s license seat. Then delete the program (Windows: and its PATH entry). In that order, Claude Code shows no hook errors afterwards. Versions before 1.0.3: tokenmaestro guard uninstall, then tokenmaestro deactivate.