Where does the /tokenkarma command come from?

The tokenkarma desktop app installs it, opt-in, when you connect Claude Code. The guard runs locally on your machine: the instant commands cost zero tokens and answer immediately, without calling the model. Uninstalling the app removes everything.

Does it work on Windows?

Same command, same guard. The Windows app is on its way, and the command grammar, the protection panel and the caps behave identically. Join the list on the Windows page to get one email the day it ships.

What does tokenkarma see?

The guard counts your Claude Code spend locally, from your own transcripts, to enforce the caps you set. Connected sources show live usage in your island and dashboard; anything beyond your plan’s display limit shows as locked until you unlock it. Credentials are encrypted at rest and revocable in one click; the privacy policy lists exactly what is collected.

July 7: Fable 5 leaves the included plans.

Cap this session's extra usage. One command.

Fable 5 goes 100% extra usage on July 7. Anthropic caps your spending once a month. tokenkarma caps it per session, per day, per week, and blocks the prompt before it spends.

Get started

tokenkarma, your token's guardian.

claude

> /tokenkarma 20

Session capped at $20.00 of extra usage (counted from now)

┌─ 🛡 tokenkarma ───────────────────────────────
│ 🔴 EXTRA USAGE — limits armed
│ /tokenkarma session|day|week <usd>
│
│ Session   $0.38 / $20.00
│ █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░   2%
│ Day       no limit
│ Week      no limit
│ Extra usage this month   178.26 / 500.00 EUR
│ ███████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  36%
│
│ ⛔ 1 block today
└─ ⚡ /tokenkarma 20 · /tokenkarma menu ─────────

The daily gesture

You start a project, you type /tokenkarma 20, you work. That is the whole workflow. This session now has a $20 budget of extra usage, counted from the moment you set it. If the session burns through it, the next prompt is blocked before it spends a cent.

Blocking is not a prison. The block message says "Re-send the same prompt to continue": send the same prompt again within 5 minutes and it goes through, because repeating yourself is explicit consent (you then get a 30-minute grace window). Walk away instead, and the money you did not spend is measured and credited: that is the "1 block today" line in the panel.

Hard enough to stop accidents, soft enough to never block a deliberate decision.

Command reference

Instant commands are answered locally by the guard: zero tokens, immediate, no model call.

/tokenkarma Your protection at a glance: caps, live extra-usage burn, blocks.

/tokenkarma 20 Cap this session at $20 of extra usage, counted from now. The daily gesture.

/tokenkarma session|day|week|month <usd> Set a cap per window. Session covers the current session; the others are global.

/tokenkarma session off Lift one cap (same for day, week, month).

/tokenkarma limit The detailed table of your caps.

/tokenkarma limit off Lift every cap.

/tokenkarma status Today's Claude Code spend and the session profile.

/tokenkarma guard off|on|eco Pause or resume blocking for this session, or switch to eco mode.

One interactive command, /tokenkarma menu (a few seconds), opens the full menu: guided caps, cheaper-model tips on or off, auto-switch on or off, and "How much did I save?".

What it protects, exactly

The caps act only while your account is in extra usage. Usage included in your Claude plan is never counted against a cap and never blocked: if you are in plan, the guard does nothing at all.

A session cap counts from the moment you set it, never retroactively. Type /tokenkarma 20 after an expensive morning and the $20 budget starts at zero, not at what the morning already burned.

Both of those rules are deliberate. A guard you cannot predict is a guard you turn off.

It also saves while you work

Cheaper-model tips

Get a tip when a cheaper model can do the job, with the estimated saving for that prompt. Demanding work is never downgraded, and micro-prompts stay silent. Follow a tip and the saving joins your measured total; ignore it and nothing is counted.

Auto-switch

Simple background tasks run on a cheaper model, automatically. You save without doing anything. Every dollar it credits is measured from the real tokens consumed, never estimated, and measured and estimated numbers are never added together.

Anthropic caps extra usage per month.
Expert caps it per session, day and week.

Every plan sees the protection panel and the live extra-usage burn. The caps themselves are an Expert feature; on Free and Hobbyist they show as locked. Prices are on the pricing page, served live.

See pricing

Frequently asked questions

Where does the /tokenkarma command come from?: The tokenkarma desktop app installs it, opt-in, when you connect Claude Code. The guard runs locally on your machine: the instant commands cost zero tokens and answer immediately, without calling the model. Uninstalling the app removes everything.
Does it work on Windows?: Same command, same guard. The Windows app is on its way, and the command grammar, the protection panel and the caps behave identically. Join the list on the Windows page to get one email the day it ships.
What does tokenkarma see?: The guard counts your Claude Code spend locally, from your own transcripts, to enforce the caps you set. Connected sources show live usage in your island and dashboard; anything beyond your plan’s display limit shows as locked until you unlock it. Credentials are encrypted at rest and revocable in one click; the privacy policy lists exactly what is collected.