SaveMyTokens

Decide which projects get your Claude allowance.

A capacity scheduler for Claude Code · Node 18.17+ · nothing leaves your machine

You have four Claude Code windows open. One of them is a throwaway experiment and it is quietly eating the afternoon, but there is one usage number for all four, so there is no way to know which. By the time the window is gone, the work that mattered is the work that did not finish.

Get started

npx savemytokens

That is the whole install. On the first run it offers to add four hooks and a status line to Claude Code's settings.json, backing the file up first.

• Say yes. The status line is the only place your 5h and weekly usage is published, and it is what proves a session is still open. Nothing is live without it.
• Don't want to see one? Take Install, no line. Everything is installed and the line renders nothing, so the numbers still arrive and your terminal looks unchanged.
• Rather not fetch it each time? npm i -g savemytokens, then the command is just savemytokens.
• Changed your mind? npx savemytokens uninstall takes every entry back out.

The real numbers	Your 5-hour and weekly usage exactly as Claude Code publishes it to the status line. Never an invented "% remaining".
A split you control	Give each project a target and a priority. Move one and the others move to fit, so the window always adds up.
The agent working to it	As a session eats into its slice it is told to narrow scope and finish, and to write down whatever it drops. That comes back the next time you open the project.
Capacity returned	A project with nothing running lends its share to the rest and reclaims it when you come back.
In your status line	⣿⣿⣿⣿⣿⣀⣀⣀⣀⣀⣀⣀ 42% of 50% · 5h 42% · resets in 3h52m. Pick a shape, arrange the pieces yourself, or run it silent.

A worked example

Seven projects open, and the release matters more than the experiment.

npx savemytokens share burningdemand 20      # Set burningdemand target to 20%
npx savemytokens priority burningdemand high # burningdemand now takes spare capacity first

The others split what is left. When one of them finishes, whatever it did not spend goes to burningdemand, because it is the only project on the high tier.

coldverdict is at 93% of its slice, so its next turn opens with a line telling it to finish what it has and write down the rest. Whatever it drops comes back the next time you open that project.

The three pinned rows hold the order you pinned them in, whatever happens to the numbers beside them. Nothing here required you to guess how many tokens anything costs.

Commands

--codex points any of these at Codex instead, which publishes its own rate limits and needs no hooks. It is metered and shown, but nothing can be injected into it, so there is no advice.

In a row, ★ means you pinned it and pinned rows keep the order you pinned them in; a number beside it is how many Claude Code windows are open in that project, shown only when there is more than one; a dimmed row is holding its share while nothing runs there.

Inside the control centre, ? lists every key. The ones worth knowing: ←→ moves an allocation, and the others move to fit so the window always adds up. space moves a project up a level and x moves it down, so one you never want to see goes ACTIVE to RECENT to hidden, and m shows what is hidden. e puts everything back to an even split, p cycles priority, ⏎ opens a project's sessions, s opens settings.

Where the numbers come from

Three different kinds of number sit on that screen, and it matters which is which.

5h and 7d	Published figures, not ours. Claude Code writes them to the status line, so SaveMyTokens installs one, reads them, and stores each reading with its timestamp. Nothing is estimated.
share	Measured from the transcripts already on your disk, token by token, including subagents. Exact, and independent of anything the provider reports.
used, allocation	The published percentage, split by that measured share. The total is not ours. The division between projects is ours, and it is the only inferred number on the screen.

Usage from another machine, or from claude.ai, is invisible to a local tool. When the window moves across five minutes in which nothing local was metered, it is reported on its own line rather than folded silently into a project's figure.

How it works

• No daemon. Three scripts in ~/.savemytokens/hooks, four hook entries and a status line in Claude Code's settings.json, which is backed up first.
• The hooks meter their own session as it runs. The status line captures the published window every ten seconds. The control centre reads what they leave behind.
• Transcripts are read incrementally, from a stored byte offset per file, so a prompt costs milliseconds however long the session has run.
• Nothing else is touched. uninstall puts it back, and restores a status line of your own if it wrapped one.

Requirements

• Node 18.17 or newer, and Claude Code 2.1 or newer.
• A Claude subscription for the published 5h and 7d figures. On an API key or through the Console no window is reported; SaveMyTokens says so rather than guessing, and allocation and advice still work from measured usage alone.
• macOS is what it is developed and tested on. Linux should behave identically. Windows is written for but unproven: paths and the home directory resolve portably, and the control centre wants a terminal that handles the alternate screen buffer, which Windows Terminal does. If it misbehaves there, please open an issue.

npx savemytokens theme new mine nord   # copy one to start from
npx savemytokens theme check mine      # is it readable?
npx savemytokens theme tui mine        # use it

~/.savemytokens/themes/mine.json. Everything you leave out is inherited:

{
  "colors": { "fg": "#eceff4", "dim": "#7b88a1", "accent": "#88c0d0",
              "ok": "#a3be8c", "warn": "#ebcb8b", "danger": "#e0707c", "track": "#434c5e" },
  "tui":    { "cursor": "❯", "pin": "★", "active": "●", "done": "✓",
              "fill": "▰", "empty": "▱", "over": "▶", "meter": "▰", "track": "▱" },
  "glyphs": { "sep": "·" }
}

theme check measures every colour against the terminal and fails anything unreadable, against the same bar the built-in themes have to clear.

reads	~/.claude/projects/**/*.jsonl, the transcripts Claude Code already writes, from a stored byte offset, and the status line payload Claude Code hands it on stdin.
writes	~/.savemytokens only: meter offsets, per-project settings, quota readings, deferred notes.
injects	A line of context on SessionStart and UserPromptSubmit when a project is over its share. Nothing else.
never	Makes a network call, blocks a prompt, or exits non-zero.

What it will not do

• It advises; it does not enforce. The hooks inject text, and a model does not hold a budget the way a scheduler holds a lock. A session that ignores the advice keeps running. Hard caps are the next step, and will ship only after a period of logging what would have been blocked.
• It does not predict a lockout. A percentage and a reset time are published. Anything beyond those two facts would be a guess dressed up as a number.
• It cannot see usage it did not measure. Another machine, claude.ai, or work done before you installed it shows up as unattributed window, labelled as such.
• The split moves under you. A window that has taken no turn in ten minutes lends its share to the rest, and takes it back the moment you type in it again. What is on screen is the current split, not a contract: a project's allocation can change without you touching anything.
• It makes no network call. No account, no telemetry, no daemon. savemytokens privacy prints every file it reads and writes.