Get more out of every Claude Code session.

1. 1

   Measured, not read

   Neither quota is written to disk. The five-hour window is chained off turn timestamps and floored to the half hour; the week is pinned to the reset Claude Code cached. If Claude has refused a turn, its own reset time wins.

2. 2

   Used is what you were billed for

   Input, output and newly-cached tokens, across every project and every subagent. Cache reads are shown apart — they cost a fraction as much and would swamp the rest fifty to one.

3. 3

   Projected looks forward

   Where the window lands by its reset at the rate it has kept. It stays hidden until a fifth of the window has gone, since an early rate is noise.

4. 4

   A yardstick, not a quota

   The bar is a share of your heaviest closed window — 7 days back for the five-hour card, 28 for the weekly one. Past it the bar reads full and the percentage keeps counting.

Transcripts are append-only, so files untouched since the cutoff are skipped on mtime and the rest are cached per version. On the development machine, 901 files and 1 GB come to ~1.3s cold and ~7ms warm. The page reads every 15 seconds and ticks the countdown itself in between.

1. 1

   One line, when the projection crosses

   The same fact the card already shows in colour, addressed to whoever is not looking at the card. It carries the reset time, because that is the half you can act on.

2. 2

   A switch each, both start off

   Settings is the third tab, with one switch per limit. Turning one on is what asks the browser for permission — nothing asks on load. If the browser is blocking the page, the switches come back off and say so.

3. 3

   Then quiet for as long as you say

   It speaks once, then holds its tongue: an hour for the five-hour window, four for the week. Crossing the line again inside that interval is the same interruption twice, so it is not sent.

4. 4

   Offered once, kept in this browser

   A sheet offers these the first time there is a limit to show. Not now means not now — it returns once, a fortnight later, and that is the last of it. Settings live in this browser, like the theme.

Nothing runs in the background: the page notices on the same 15-second read the cards already make, so a tab has to be open. And a week with no reset from Claude has nothing to project into, so it never speaks.

1. 1

   Spend per day, quiet days included

   A bar for every local day in the range, quiet days included. Heights are linear, so one heavy day flattens the rest. A mark under a bar is a day Claude refused a turn.

2. 2

   Hour of day, a month folded onto one week

   Every half hour of the range laid over one week, darker where it is heavier. Tells you what the daily bars cannot: whether the window opens at nine in the morning or eleven at night.

3. 3

   Every project and model, ranked by what it billed

   Cache reads held apart, as on the limit cards. Pick a project and the summary, both charts and the model list narrow to it; pick it again to let go.

4. 4

   The range and the project live in the URL

   `?range=7d&project=…`, so reload, bookmark and Back all work. Ranges are whole local days, capped at 90.

• Both limits, at the top

  How much of your five-hour and weekly windows you have spent, when they reset, and where they land at this rate. Subagents included.

• A heads-up before you hit one

  A desktop notification when a window is headed past your heaviest one. One switch per limit, both off until you turn them on.

• Where the tokens went

  A second page: spend per day, per half hour, and per project. Pick a project and the whole page narrows to it.

• What the spend was made of

  Oversized sessions, standing context, a lopsided model mix — measured against your own machine's middle, so an ordinary week says nothing.

• Active sessions, verified twice

  Every running session is checked against the OS, so a row that says working really is one to wait on.

• Recent sessions, every project

  Title, first and last prompt, model, and branch — enough to find the one worth resuming. Filter by day or a range of your own.

• A panel per session

  Message and tool-call counts, tokens, elapsed against working time, subagents, a copyable resume command, and the transcript.

• The standing cost, itemised

  What the fixed half of a session's window went on — CLAUDE.md, skills, deferred tools, agents, MCP — priced block by block.

• Token usage you can sort by

  Tokens per session, coloured by share of the context window, so you see what is about to compact. Ranked across the whole window.

• Scriptable output

  A --json flag prints the same payload the page uses. The HTTP API behind it is open too.

• Fully keyboard driven

  Filter, move between sessions, open and close the panel. A check takes seconds.

• No dependencies, no install scripts

  Nothing to install, nothing to configure, and the same behaviour under npm, pnpm, yarn, and bun.

• Local by construction

  Binds to loopback, refuses non-loopback hosts, makes no outbound calls, and never writes to your Claude directory.

• Never writes to your Claude directory

  It reads what Claude Code already put on disk, opens nothing for writing, and creates nothing.

• Loopback only

  Binds to 127.0.0.1 and refuses any request whose Host is not loopback, so no website can point your browser at it.

• No outbound calls

  No telemetry, no update check, no analytics. Nothing leaves the machine.

• No dependencies, no install scripts

  Nothing third-party runs on your machine, at install time or after.