Make your AI coding plan do more for you. From di first request reach di last token.
Dem dey build TokenMaestro to conduct di work between you, your AI agent and your computer: one local model go structure your request, ask wetin dey miss before di agent start, hand heavy work to tools wey dey installed for your machine, and give each task di right model and effort. Every saving, dem go measure am for your machine, e no go be estimate.
First release, for Claude Code: di free X-ray and, inside Pro, di session guard. Next: di full orchestra, plus Codex, Gemini and oda agents.
Free X-ray · No account · Nothing dey comot from your computer
Di orchestra: less tokens from di start
Most token tools dey compress wetin di agent don already spend. TokenMaestro dey work before di spending: for di request, di understanding, di tools and di model.
Request wey dem build well
One small open model dey run for your computer and e dey turn wetin you type into clear, structured request before e reach di agent. E no dey cost any token.
Questions before di work
When request no clear, TokenMaestro dey ask you first, so di agent no go spend tokens to dey guess and dey redo work.
Local tools for di heavy work
To read text inside images, convert documents, cut video: tools wey dey installed for your machine dey do am, and TokenMaestro dey set dem up with your permission. Di agent only dey get di result.
Di right model and effort
Simple tasks dey go to lighter model with low effort; di heavy model and high effort na only when di task need dem.
Lean sessions
Warnings inside Claude Code, each one with im cost for API prices: when di context don big well well, wetin every step wey follow dey pay to read am again; when one break don make di cache expire, wetin e cost to write am again compared to one turn wey get cache. When you resume one session, di warning dey come before your first message; after one break for middle of session, e dey come with di message wey you just send.
Every token, accounted for
Di X-ray dey show where every token go, counted from di logs, no estimate. Next: dem go measure each fix on top your own work, on and off; if e no pay, dem go switch am off.
Your agent, your choice: Claude Code first; Codex, Gemini and oda ones next. TokenMaestro dey conduct; na you dey choose di instruments.
Why dis one dey save tokens
Inside AI agent, every step dey re-read everything wey come before am. One token wey remain for di start, you go pay for am again for every step after, and one token wey dem cut from di source dey save for all of dem. Na why TokenMaestro dey work before di spending.
Clear request, less loops
Request wey no clear dey turn to try, fix, and try again, and every loop dey re-read di whole session. To structure di request first dey cut di loops.
Local question no cost anything
To clear doubt for your computer no dey spend cloud tokens. If di agent come find out say e no understand well, e go cost one full round, plus di fix.
Heavy files dey stay outside di context
One 50-page PDF wey di agent read go enter di context, and e go dey re-read am for every step. One local tool go only comot di part wey matter.
Model wey fit di task
Light model dey cost small part of wetin di strongest one cost, and high effort dey produce plenty reasoning text. Di strong model go only enter when di task need am.
Session wey no dey swell
Once di context pass some hundred thousand tokens, every answer go dey cost plenty. To compact for di right time, and no allow di cache expire during break, dey stop you from rewriting everything for full price.
Measured, no be promise
Dem go measure every adjustment on top your own work, on and off. If e no save, dem go switch am off. Until then, nobody dey promise any saving.
How e dey run, from start to finish E dey come
- You go install am one time. TokenMaestro go find di agents wey you dey already use and di tools for your machine.
- You go make your request through TokenMaestro: for di terminal, or from inside di agent itself.
- E go prepare di request for your computer (structure am, ask wetin dey miss, run local tools) and hand am over, ready, to di agent wey you choose.
- During di session, e dey watch context, cache and limits, and e go warn you before dem turn to waste.
- For di end, di X-ray go measure wetin change.
Today: di X-ray dey read Claude Code logs whether you sign in with subscription (Claude Pro or Max) or with API key, and e dey show di value for API prices for both. Next: di orchestra go drive di official apps of di agent wey you choose (Claude Code first; Codex and Gemini next), di same way wey you dey sign in to dem.
Start with di X-ray: how much your plan worth, and where e go
One command dey read months of logs inside seconds and e go write one dashboard file: value for API prices by day, model, effort level and project; di days wey you hit di limit; and cross-check against Claude Code own counters.
Wetin e dey find: di leaks, with label for every number
“Measured” mean say di logs show am. “Ceiling” mean di most wey one fix fit recover — e no be saving. For di founder logs:
Very big contexts
of di value for API prices go to turns wey get more than 400k tokens of context. Na share of di spending, no be money wey you fit get back. Every step dey re-read everything wey come before am.
Cache wey dem rewrite after breaks
Di cache expire 491 times during pauses, and dem write di part of di context wey no dey inside di cache again for full price.
Tokens before di first answer
Instructions, tools and memory wey every session dey start with — dem dey re-read am for every turn after dat. Dem dey count am up to 100k tokens per session.
Di weight map
Wetin your agent read di most, by project and file. Di area na di volume wey enter di context; di colour dey show how much of am dem re-read when di file never change.
Next: di same map, drawn on top your code structure, with di path wey your agent take for each session.

Why you fit trust di numbers
Most token tools dey report savings wey dem estimate by demsef. Independent measurements of di popular ones no find any savings, and sometimes dem find higher cost. We start from di oda side.
Di log traps, handled
Claude Code dey write plenty records for one message, and di early ones get partial counts; sessions wey dem resume dey replay messages; subagents dey write separate files. Each case get test.
Checked against Claude Code
For 16 comparisons, our cache-read count come out 4.6% below Claude Code own counters, and e never pass dem. Di numbers na floor.
No savings claims until we measure
No savings percentage go enter dis page until dem publish one paired measurement with di raw data and one way to reproduce am.
Download
Di X-ray na small command-line tool. Paste one line for your terminal to install am, or download di file.
irm https://thetokenmaestro.com/install.ps1 | iexWindows SmartScreen fit show warning for di file wey you download, because di program never get signature yet. Di install line wey dey up na di way wey we recommend.
curl -fsSL https://thetokenmaestro.com/install.sh | shtokenmaestro xrayYou dey work inside Claude Code? You fit ask am make e run di install line for you.
To update, run di install line again. To uninstall: tokenmaestro guard uninstall (if di session guard dey on), then tokenmaestro deactivate (if you get Pro license), and na after dat you go delete di program (for Windows: di PATH entry too).
TokenMaestro dey run for computer: Windows, macOS or Linux. Open dis page for your computer to download am.
Price
X-ray
- Value and usage dashboard
- Leaks, with evidence labels
- Weight map
- Cross-check with Claude Code counters
- Check for tools wey dey miss
Pro
Founding members: US$19 a year, and di price no go change as long as you dey subscribed. First 1,000 pipo.
- Session guard: warning inside Claude Code when di context pass 200 thousand tokens, with wetin e cost to read am again for every step wey follow, for API prices
- Warning when one break don make di cache expire, with di extra cost to write di context again compared to one turn wey get cache
- Status line for di terminal: context, re-read cost of di next step, cache and plan limits
Di status line dey work inside di Claude Code terminal; tokenmaestro guard install go show you how to on am.
- Structured requests from one local model
- Questions before di agent start
- Local tools, installed and tuned
- Di right model and effort for each task
- Savings wey dem measure on top your own work
- Codex, Gemini and oda agents as dem dey come
Di license key go come by email. Activate am with tokenmaestro activate YOUR-KEY. E dey work for up to 3 computers. Di session guard need Claude Code 2.1.251 or newer.
Two years of Pro cost less than one single month of di difference between one Claude plan and di next one.
Questions
Wetin TokenMaestro dey read?
Di session logs wey Claude Code don already dey write for your machine (~/.claude/projects): di token counts wey di API return for each message, and which tools run and how much dem return. Nothing dey go anywhere.
Na di dollar value be wetin I pay?
No. Na wetin your usage for cost for API prices. For flat-rate plan, e dey show how much your plan deliver, e no be bill.
Claude Code already get /usage. Why dis one?
/usage dey tell you how much of your limit don go. TokenMaestro dey tell you where e go, how much e worth, and wetin dey leak.
How much I go save?
We no dey promise any number before we measure. Di X-ray dey show how much dey leak for your case. When di fixes come out, dem go measure each one on top your own work.
I need account?
No. Di X-ray no need account or email. Pro dey use license key, wey dem go send by email when you buy.
E dey work with Codex, Cursor or Gemini?
Not yet. Today e dey read only Claude Code; Codex and Gemini go come next, and Cursor no get support for now. You go choose di agent; TokenMaestro go conduct di rest.
Windows, macOS, Linux?
Yes. Di X-ray dey run for all three.
I need to install models or tools?
No. When one task need am, TokenMaestro go show wetin dey miss, explain why, and install am with your permission. Di local model small and e dey open; if your computer no fit run am well, TokenMaestro go work without am. (Pro, e dey come.)
E dey work with subscription, or only with API key?
Both. Di X-ray dey read di logs wey Claude Code write, no matter how you sign in, subscription (Claude Pro or Max) or API key, and e dey show di value for API prices for both. E no fit show how much of your plan limit each thing use, because dem no publish dose limits; but e dey mark di days wey you reach di limit. Di orchestra, when e land, go drive di official apps di way wey you dey sign in to dem, and e no go ever use your login outside dose apps.
Refund dey?
Yes. You fit ask for full refund within 7 days of di purchase if you dey live for Brazil, within 14 days if you dey live for EU, EEA or UK, and anywhere else within 7 days of your first purchase of one plan. Write to hello@thetokenmaestro.com. Creem, our merchant of record, go process di refund, and after dat dem go disable di license key. Di X-ray na free, so you fit see wetin TokenMaestro find for your own machine before you buy.
How I go take uninstall am?
First run tokenmaestro uninstall: e go remove di session guard hooks from Claude Code and free di license seat for dis computer. Then delete di program (for Windows: di PATH entry too). If you do am for dis order, Claude Code no go show hook errors after. Versions before 1.0.3: tokenmaestro guard uninstall, then tokenmaestro deactivate.