Illustration: a small creek with soft trees. CTO's platypus sits at a tiny wooden desk on a flat stone, typing on a laptop. Across the water, an empty hammock with a book on it hangs between two trees.

Free token coach for Claude Code

Work on what matters.
CTO keeps output quality up and token usage down.

Quiet in the background. One line, with the fix, only when something is worth it.

claude plugin install --marketplace tokencoach/token-coach token-coach

Paste in a terminal, then fully quit and reopen Claude Code.

Runs locallyFree to startAbout 500 tokens a sessionWorks with Codex and Cursor

Sound familiar?

Claude Code is worth it. Some of your habits may not be.

usage your task

Limit hit mid-task

No telling which habit used it up.

turn 1 turn 2 turn 3 turn 4 turn 5

Every turn gets slower

A long thread re-sends everything, every turn.

limit average five-hour windows, one week

Your plan may not fit

Limits bind on your busiest five hours, not your average.

See it work

One session with CTO switched on.

  1. install$ claude plugin install --marketplace tokencoach/token-coach token-coach
  2. small task> rename the variables in utils.py to snake_case
  3. No CTO line. At this size, nothing is worth doing but the task.
  4. risky step> drop the users table
  5. CTO: This task includes a step that can't be undone ("drop the users table"). Which step here is irreversible, and where do you want to approve before it runs?
  6. long thread> …two hours later
  7. CTO: This thread has done a lot of work, and the context window is 86% full (172k of 200k). Want me to write the handoff note now? Then /clear is yours to run.

How it works

Quiet until it matters. Then one line, with the fix.

  1. Small task

    No CTO line.

  2. A step that can't be undone

    CTO: …where do you want to approve before it runs?

  3. Context filling up

    CTO: …86% full (172k of 200k). Want me to write the handoff note now?

  4. “Did it pay off?”

    Graded against your own sessions, with receipts.

Shortened from CTO's real messages. Claude passes them on in its own words, so the wording varies.

Beside /usage and /insights

Claude Code shows what you spent. CTO adds what to do about it.

/usage and /insights

  • What you spent
  • General tips

Good. Use them.

CTO adds

  • FitWhat's worth it for this task, and what to skip
  • Pay lessWhen your plan is bigger than your usage
  • ProofGrades its own advice against your later sessions
  • PauseAsks before a step that can't be undone

Our own numbers

Light to run. Graded on your own sessions.

Every recommendation is checked afterwards against your later sessions, and counted only if the numbers moved.

40
techniques it knows

Each with the task size where it pays, and where it doesn't.

~500
tokens of overhead a session

As Claude Code's own claude plugin details measures it.

$0.02
CTO's own cost over 30 days

Against $1,966 of Claude Code usage in the same 30 days.

CTO's platypus with a small shrug: not yet.

Privacy

What leaves your machine

  • Your transcripts, prompts and code: never.All analysis runs locally.
  • Technique and price updatesOnce a day. Sends nothing about you. You can turn it off.
  • The weekly emailOnly if you set it up, to your own inbox.
  • The outcome reportNever sent by CTO. You decide whether to share it.
  • The Pro connectorKeeps your sign-in email and subscription status. What you ask it is answered, then not stored.

Install

One command. No setup.

Needs Claude Code with plugin support and python3. Nothing to pip install.

  1. Install CTO:

    claude plugin install --marketplace tokencoach/token-coach token-coach

  2. Fully quit and relaunch Claude Code (a new thread is not enough). CTO is on from your next session.

On Claude Code older than 2.1.292? Use two commands.
claude plugin marketplace add tokencoach/token-coach
claude plugin install token-coach@token-coach-marketplace

Want to know first whether you'd gain anything?

One file, no install. It reads your local history, prints a verdict (including "already efficient"), and makes no network calls of its own.

curl -fsSL https://raw.githubusercontent.com/tokencoach/token-coach/main/tools/qualify.py -o /tmp/q.py && python3 /tmp/q.py

Using Codex or Cursor? The task check, the prompt check and the pause before irreversible steps work there too: how.

Pricing

Free to measure. Pro to coach.

Free

$0

The whole plugin, inside Claude Code.

  • Spend, trends and where the money went
  • The plan check, including when to pay less
  • The task check before a piece of work: what's worth doing, what to skip, which model
  • The prompt check: what your prompt is missing
  • The pause before a step that can't be undone
  • Context and cache alerts while you work
  • "Did it pay off?", graded against your own sessions
  • Learns what works for you, per repo
  • The weekly digest, handoff notes and the status line
Install free

Pro Opening soon

$19 / month

CTO everywhere you work, not just Claude Code.

  • The CTO connector: sign in from Claude, ChatGPT, Codex or Cursor
  • The task and prompt checks in each of them
  • Cancel any time; refund within 14 days, no questions

Not open yet. Join the list and we'll write once, when it opens.

Join the waitlist

Team Coming

Pro on every seat, with a team roll-up and admin. Not built yet, so it is not for sale.

If you want it, tell us and we'll write when it ships.

Tell us you want it

Everything in the plugin is free, for good. Ask CTO "did it pay off?" and it grades its own advice against your sessions.

Roadmap

What is coming (not built yet)

Listed so nothing is hidden.

  • Published savings receipts from real users, as they come in.
  • A quality measure: did the advice make results better, not just cheaper.
  • Cross-tool usage (Cursor, Copilot) beside Claude Code.
  • Better matching of tasks described in your own words, tested on a larger held-out set.
  • Router mode for API-key users, being evaluated.
  • Team dashboards and the Team plan.

FAQ

Questions

How do I reduce Claude Code token usage?

Install CTO. Before a task it says which techniques are worth it at that size and which to skip, it warns when a long thread is re-sending more than it should, and it offers a handoff note so the next part starts clean.

Why do I keep hitting the Claude Code usage limit?

Limits bind on your heaviest five hours, not your monthly average. CTO shows which habits use your window and what to change, so you get more done before the limit.

Which Claude plan should I be on?

Run the plan check. CTO compares what you use, and your busiest five-hour window, with what you pay, and tells you to keep, switch or pay less.

Does my code or my prompts leave my machine?

No. All analysis is local. CTO downloads technique and price updates once a day and sends nothing about you or your usage.

How is CTO different from Claude Code's /usage and /insights?

Those show what you spent and suggest general improvements. CTO adds four things: which techniques are worth it for the task in front of you, advice to pay less when your plan is bigger than your usage, grading of its own advice against your later sessions, and a pause before irreversible steps.

Will CTO slow Claude Code down or fill my context?

Its always-on cost is about 500 tokens a session, as Claude Code's own claude plugin details measures it. Its hooks cost nothing until they speak, and they speak only when something is worth it.

Does CTO save money?

That's its job. It flags which techniques cut tokens for the task in front of you and which to skip, warns before the context window fills, and tells you when your plan is bigger than your usage. It then grades its own advice against your later sessions, so you can see what worked for you.

Does it work with Codex or Cursor?

Partly. The task check, the prompt check and the pause before irreversible steps work in any MCP client or through an AGENTS.md pack. Spend, plans and proof need Claude Code's own logs.

How do I uninstall it?

Inside Claude Code, run /token-coach:settings uninstall: it shows what CTO added outside its own folder and removes it when you say yes. Then run claude plugin uninstall token-coach@token-coach-marketplace.

Is CTO made by Anthropic?

No. CTO is an independent plugin that works with Claude Code. It is not made by, affiliated with, or endorsed by Anthropic.

Why a platypus?

It hunts with its eyes shut, sensing what it can't see. That's the job: noticing what your usage is doing, so you don't have to watch it.