claude-usage-tracker 0.1.18

Windows tray widget + dashboard for Claude plan usage (5-hour & weekly limits)

Claude Usage Tracker

A small Windows desktop widget that tracks your Claude plan limits — the 5-hour and weekly usage windows, the exact numbers the /usage command shows — with a live always-on-top widget, a glass dashboard, a tray icon, and a toast notification every time usage crosses a 20% mark.

It reads the OAuth token Claude Code already stores on your machine and calls the same endpoint /usage uses. Read-only — it never modifies your credentials and talks to nothing but that one Anthropic endpoint.

The always-on-top mini widget (sample data)

The full dashboard — gauges, live reset countdowns, burn-rate, history, overage credits, and Open Sessions (sample data)

Open Sessions — per-project context-window fill % (toggle to Tokens), with active indicators (sample data)

Features

• Always-on-top overlays — a frameless, draggable, resizable mini widget (5h, weekly, context %) with Refresh / Check-for-updates buttons and a dropdown to pick which session's context to track; and a minimal HUD bar — a translucent, one-line, FPS-counter-style strip (dir Ctx: 45% 5h: 11% 7d: 3%) whose controls appear on hover. Toggle either from the tray; they can be shown together. Drag a corner to resize; size is remembered.
• Full dashboard — animated ring gauges, live reset countdowns, a burn-rate / time-to-limit projection ("≈7%/h · hits 100% in ~1h 10m"), a usage history sparkline, overage credits, and per-model (Opus/Sonnet) scoped weekly limits.
• Open Sessions — per-project context-window fill % and last-5h token usage, read from your local Claude Code logs (token counts only, never content), with Context % / Tokens tabs and active indicators — see which terminal is burning your usage.
• All-time stats — a second dashboard tab that mines your whole local history: an Overview (sessions, messages, total tokens, active days, current/longest streak, peak hour, favorite model, a contribution heatmap, and a "you've burned N× more tokens than War and Peace" line) and a Models view (tokens-over-time stacked chart + per-model input/output split), all with a 7d / 30d / All toggle. Folded incrementally from your local logs (each file read once, then only new bytes), so it's accurate without rescanning gigabytes — token counts only, never content.
• Proactive alerts + traffic-light verdict — toasts when a window is on track to run out before it resets, when the active context hits 90% (time to /compact), or when overage credits near the cap; plus a one-glance green / amber / red verdict in the widget and dashboard, and a daily update check with a one-click in-app Update that upgrades in place and restarts — the .exe runs the signed installer, and pip/pipx installs upgrade themselves via the app's own Python (no pipx on PATH needed, no trip to GitHub). Refresh and Check for updates are one click away in the dashboard, widget, and tray menu too.
• Live tray icon — two bars (left = 5h, right = weekly) that fill and change colour with usage.
• 20% notifications — a Windows toast each time the 5h or weekly window crosses 20 / 40 / 60 / 80 / 100%. The first reading is recorded silently, so you only get pinged on future crossings, never a burst at startup.
• One-click sign-in — when your login expires, a Sign in to Claude action (in the tray menu and on the dashboard banner) runs claude auth login — Claude Code's own sign-in — so you can refresh it without opening a terminal. The app never writes your credentials itself; it just triggers Claude Code's flow.
• Auto-start on login, single-instance, graceful rate-limit (HTTP 429) back-off, and automatic pickup of account/token changes (it re-reads your login each poll).

Colour scale

Bars and gauges share one scale (low → high):

Range	0–20%	20–60%	60–80%	80–90%	90–99%	100%
Colour	green	blue	orange	red	black	yellow

The 90–99% "black" band gets a red glow/outline so it reads as danger rather than empty.

Tray icon at 5/40, 45/72, 85/94, 95/100, and 0/0 percent

Install

Requires Windows 10/11, Claude Code installed and logged in (so ~/.claude/.credentials.json exists), and the Edge WebView2 runtime (preinstalled on Windows 11; otherwise a free download). Then pick one:

Option A — Installer (easiest, no Python needed)

Download ClaudeUsageTracker-Setup.exe from the latest release and run it. A standard Windows wizard: accept the license, then tick the shortcuts you want — Desktop / Start Menu / start at sign-in.

Windows SmartScreen ("Windows protected your PC") appears because the installer isn't code-signed — normal for small open-source apps. Click More info → Run anyway.

Option B — pipx (updatable from the command line)

pipx install claude-usage-tracker
claude-usage-tracker --install      # interactive: Desktop / Start Menu / Startup
pipx upgrade claude-usage-tracker   # later, to update

Option C — from source

git clone https://github.com/paris-paraskevas/claude-usage-tracker.git
cd claude-usage-tracker
python install.py

After any of these, the mini widget appears top-right and a tray icon by the clock. To pin it to the taskbar (Windows blocks apps from doing this themselves): right-click the running app's taskbar icon → Pin to taskbar.

Updating: the app checks daily and shows an Update to vX item in the tray. For the installer build it downloads and runs the new Setup.exe for you (closes, upgrades, relaunches); pipx installs use pipx upgrade claude-usage-tracker.

Usage

Launch "Claude Usage Tracker" from the Start Menu/Desktop, or run it directly:

.venv\Scripts\pythonw.exe claude_usage_tracker.py

Tray menu: Show/Hide widget, Open dashboard (native window), Open in browser, Refresh now, open config/log, Quit. The widget has a × to hide it; drag it anywhere.

CLI:

.venv\Scripts\python.exe claude_usage_tracker.py --once          # print status once
.venv\Scripts\python.exe claude_usage_tracker.py --once --debug  # + raw API JSON
.venv\Scripts\pythonw.exe claude_usage_tracker.py --widget       # just the widget
.venv\Scripts\pythonw.exe claude_usage_tracker.py --window       # just the dashboard window
.venv\Scripts\python.exe claude_usage_tracker.py --uninstall-autostart

Configuration

Edit config.json (created on first run, in the app's data dir), then restart:

Key	Default	Meaning
poll_interval_seconds	60	How often to check usage.
threshold_step	20	Ping every N percent.
windows	["five_hour", "seven_day"]	Which limits to notify on.
notify_at_100	true	Ping when a limit hits 100%.
notify_on_start	true	One summary toast at launch.
dashboard_port	8787	Local dashboard port.
show_widget_on_start	true	Show the mini widget at launch.
widget_width / widget_height	392 / 150	Widget size in pixels.
alltime_days	30	Days shown in the All-time daily-usage chart.

How it works

GET https://api.anthropic.com/api/oauth/usage with the bearer token from ~/.claude/.credentials.json (claudeAiOauth.accessToken) and the anthropic-beta: oauth-2025-04-20 header. The response carries five_hour, seven_day, scoped per-model weekly limits, overage spend, and reset timestamps — the same data the CLI's /usage renders.

Token refresh is handled by Claude Code itself; if your login expires, the tracker shows an error state until you run any claude command to refresh it. The token is only ever read — never written, logged, or sent anywhere but that endpoint.

Privacy

Read-only and local. It reads your Claude login token and session logs from ~/.claude and displays the numbers; it sends nothing anywhere except the Anthropic usage endpoint (using your own token). No telemetry, no analytics, no data collection.

Code signing

Windows builds are signed with free code signing provided by SignPath.io, with a free code signing certificate from the SignPath Foundation.

Layout

claude_usage_tracker.py   the whole app (tray, server, dashboard+widget HTML, poller)
install.py / uninstall.py shortcut setup / teardown
requirements.txt          pystray, Pillow, winotify, pywebview
docs/                     screenshots

Runtime files (config.json, state.json, history.json, *.log) live next to the script (or in %LOCALAPPDATA%\ClaudeUsageTracker when packaged) and are git-ignored.