A native Windows tray monitor that shows your rate-limit usage as a crisp, DPI-aware icon — with burn-rate projection and a privacy-first 24h usage breakdown. Zero config: it reuses Claude Code's own login.
Built for clarity on real displays — it lives entirely in the tray and tells you exactly where you stand.
The percentage is drawn as a vector at the exact size the tray requests — no downscaled bitmaps. Stays razor-sharp on 125–200% displays. Its colour says what the number is about: white for the 5h session, yellow for the 7d week, and orange whenever extra usage is paying — which is the one thing on the tile that still says so when 0 left leaves no fill bar to.
A vertical bar rises from the bottom proportional to your usage, with a 3D bevel for relief. Blue when you're on track, red when you're not.
Projects when you'd hit 100% — warning you before the window resets, not after. A proportional pace line for the 7-day week, burn-rate regression for the 5h session.
A local, cost-weighted breakdown of your last 24h — by model, subagents, and large-context requests. Computed offline from your transcripts.
No API key to paste, no env var to set. It reuses the OAuth token Claude Code already stores when you log in.
Checks GitHub Releases and offers an in-app one-click update. Per-user install, no admin rights needed.
The tooltip packs everything into a single read — no dashboards, no browser tabs.
The icon's color tells you what's coming. The 7-day week uses a proportional pace line — your usage against the share an even burn would have spent by now — accurate from the very first reading. The 5h session uses least-squares regression on a short rolling history of your utilization.
At your current pace, usage stays under 100% until the window resets. The fill bar stays its normal blue — no action needed.
At your current pace, usage will hit 100% before the reset. The fill bar turns vivid red so you can ease off in time.
You're past the quota included in your plan and still working, because extra usage is enabled. The fill bar turns clay rather than red — red means stopped, clay means this is costing money. The tooltip says which — and since when, because "you are paying" invites "how long", and the answer is what makes the next hour of work a choice rather than something you find at the end of the month. System information tells you whether the allowance is merely enabled or actually in use. In Statistics each pane's verdict chip goes clay and reads Extra usage instead of the red At limit, which on a quota surface means stopped. What it shows you is how much of your extra-usage allowance has gone — its own window, with its own limit and its own reset — never mixed into the 0–100% of the quota included in your plan. And the number stops following your Show on icon pick for as long as it lasts: a 5-hour session with room left cannot report quota the account no longer has, so the icon reports the window that actually crossed — 0% left, or 100% used. Your pick comes back at the reset.
A window is spent and nothing is paying past it, so the account is blocked — and it stays blocked on the window that still has room. A week at 47% behind a spent session now wears the red Blocked chip in Statistics and says so in words, where it used to read On track over work that would not run. The tooltip does the same and deliberately says less: with the icon on the window that still has room, it gives no percentage and no window name — only that you are blocked until your quota resets, because naming Week 7d there would caption the wrong number.
The Statistics window draws each rate-limit window as a burn-up chart: your real utilization against the even-pace line, with a dashed projection of where your current pace lands. Two tabs — the 5-hour session and the 7-day week — each with its own verdict, plus Throughput for what is running right now and Sessions for where it all went.
The weekly projection follows the hours you actually work — flat through nights and weekends (shaded usually idle), climbing while you're at the keyboard. So “you run out here” lands mid-afternoon, not at 4 a.m. Needs a few weeks of history; until then it's the plain average-pace line.
“Stop now, pick it back up around Thursday 13:00, and you'd close the week at about 97%.” The earliest hour you normally work that would still save the week — and if there isn't one, no advice rather than invented advice.
Last week's burn-up is a faint line on the same axes. “Worse than last week?” becomes a glance instead of a memory test — no second chart, no extra number. If that week ran past the quota included in your plan, the stretch it was over is marked in clay at the 100% line, where that claim actually lives — and a week too old to have recorded it says nothing rather than pretending it stayed inside. That mark and this week's shaded stretch share one legend entry, past your plan's quota: one fact about two weeks, with the hover saying which is which.
Window average — what the window has cost so far. Now ≈ 870 tok/s — what's running this minute. Both are token throughput, not quota.
The rolling rate sampled once a second, with the ceiling ruled and labelled in tok/s. A line slides down through a pause because the work really is ageing out of the 60-second window, and climbs when new work lands — its right-hand end is the number printed above it. It scrolls because the axis is time — the motion is the data, not a spinner. Idle goes flat, and it only repaints while the window is on screen.
The upper chart is one line per repo — the heaviest four, direct-labelled with their own rates, the rest as one grey “others”. Each keeps its colour while it is on screen, so nothing changes hands mid-read; the lower chart splits the same rate by token type. Because the useful question becomes where, not how fast.
All of it comes from your real logged utilization and your local transcripts as they're written — no extra API calls. Nothing leaves your machine.
Every reset gets a celebratory, on-brand toast: a confetti burst and the quota bar visibly refilling. Each kind has its own color and headline, so you can tell what happened the instant it appears — without reading a word.
Your weekly limit reset ahead of its scheduled deadline — a known Claude Code quirk worth catching.
A partial mid-window credit dropped your weekly usage (e.g. 91% → 50%). Found money.
The calm, routine weekly reset — fresh quota for the week ahead.
The 5-hour session window rolled over — fresh for the next 5 hours.
All four are on by default — toggle each one independently in Settings.
Not every thing worth knowing is good news. These three use the same toast and say so plainly — one about what every session costs you before you type, one about the moment your quota stops being included, and one confirming a change you cannot otherwise see.
A project's instruction files, memory and skills have grown enough that every session pays for them before you type a word — with what it costs to load on a cold cache. At most once per project per week, and off by default: nobody asked to be told their own files are growing.
It exists because the four above do: the app interrupted you to say quota came back and said nothing when you started paying. Fires once when a check first finds you past your included quota — not on a timer, not again while the same spell lasts. A receipt, not a prompt: it names no account and suggests nothing.
Writing the machine-wide profile is the least visible thing this app does — nothing on screen changes, and you cannot see the effect until you start another program. So it confirms itself: only when you switch by hand, once per switch, and only after the variable is read back to check the write took. No quota bar — a bar animating one account's remaining quota into another's would suggest switching accounts because one ran out, and this app does not say that.
The Usage insights (24h) menu is computed entirely on your machine from Claude Code's session transcripts — weighted by per-model price so the numbers reflect real cost, not request count.
Claude Code loads your instruction files, your memory index and every skill's description before your first prompt — a toll paid on every request of the session, and normally invisible.
Every finding comes with one plain sentence and the concrete fix — and the only action is Copy cleanup prompt, which hands paths and numbers to Claude Code. This app measures your files; it never edits them.
AGENTS.mdSettings has a page that answers it without you opening a single JSON file — and it is read-only, like everything else here.
The rate-limit tier the API reports for your login, read as you'd say it — Claude Max 5x — with the raw tier, seat and billing type underneath. Plus whether extra usage is enabled — and, when the API is refusing it, why — and how long your stored sign-in is valid.
Claude Code keeps one account per configuration folder. Register them once and
Open Claude Code becomes a submenu — each profile launches with its own folder, its own
login and its own working directory, so the work account can't open in a personal repo. The tray
never writes to a config folder and never moves credentials: it passes
CLAUDE_CONFIG_DIR and lets Claude Code do the rest. With more than one registered, the
icon wears a thin per-profile colour band along its top edge, so a glance says whose percentage you
are looking at — in hues the icon never uses for anything else, so it can't be read as a warning.
Turn on Use the chosen profile everywhere in Windows — the switch sits in the
Profile submenu, where the picking happens — and picking a profile writes
CLAUDE_CONFIG_DIR into your user environment — so a terminal you open yourself and an
editor started from the Start menu use it too, not only what the tray launches. It follows the
profile you choose, never the one auto-follow drifts to. Open Claude Code becomes a
single command again while it is on — the profile is already picked, so every entry would open the
same one. Off by default, and reversible: switch it off and the variable, and the submenu, go back
exactly as they were. And when CLAUDE_CONFIG_DIR names a profile other than the one on
the icon, the submenu marks it · set in Windows — whose numbers you are looking at and which
profile a new session starts in are two questions, and the tray only speaks up when they differ.
Turn on Follow the active profile and the icon moves to whichever profile Claude Code last worked in — no click to switch. Read from transcript timestamps only, on the refresh you already pay for: nothing is opened, and nothing runs in between. Picking a profile by hand pins the icon there until you click Resume following. Off by default.
Account holder, email and absolute paths are hidden until you ask for them, because this is the page that ends up in a screenshot. Show reveals them, Hide puts them back.
The Statistics window reports on one profile at a time and says which. Switch it at the top and the charts, the projection, the weekly shape and the live throughput are all recomputed from that account's own readings and its own transcripts.
A folder you've never signed into runs on whatever ANTHROPIC_API_KEY is in your
environment — billed per use, and invisible to the 5-hour and weekly windows. The page says
which credentials each profile really uses, and warns in plain words when the percentages (and
the icon) aren't describing what Claude Code is spending.
Installed CLI version, install method, auto-update state, how many projects Claude Code
tracks, and the configuration folder — with a button to open it. CLAUDE_CONFIG_DIR
is honoured.
Windows build, .NET runtime, architecture and the tray's own version and data folder — the whole page as plain text in one click, exactly as shown, so a masked holder stays masked.
Only token counts, model ids, and flags are read — never your message content, with one
exception you can see: the Sessions list names each conversation, using the title Claude Code
generated for it or the prompt that opened it, because a list you cannot name is a list you cannot
search. Both are capped at 200 characters before they are stored, they appear on that one screen,
and nothing further into a conversation is ever read.
The insights scan is bounded to files touched in the last 24h and runs in the background.
The rate-limit reading is a single 1-token API call every 5 minutes. Nothing leaves your machine
beyond that, and there's nothing to log into.
Usage history is stored locally on disk and never uploaded.
The context scan reads only sizes, names, timestamps and frontmatter from ~/.claude —
never the contents of your memories, instructions or skills — and it never edits them:
the one action it offers is copying a cleanup prompt for Claude Code to act on.
Already using Claude Code? Then you're already set up.
Install it and run claude once so it stores your login token.
One command with winget, or grab ClaudeTray-Setup.exe from Releases. Per-user, no admin.
The icon appears in your tray and starts reporting. Right-click for options.
Install from the Windows Package Manager:
winget install alegauss.ClaudeCodeTray
Prefer a direct download? Grab the installer below.
