convez.ai
SkillProdutividadeIntermediário
Dubibubii

Claude Usage Widget

Dubibubii/claude-usage-widget

Monitor sempre-ao-alcance de limites de uso Claude com insights de economia

Ver no GitHub
InicialPluginScore · de 0 a 100

Widget de desktop sempre visível que monitora seus limites de uso do Claude em tempo real — janela de 5h, limite semanal, créditos Agent SDK e insights de economia de tokens. Rastreamento gratuito; personalizações estéticas custam US$5/mês.

Instalação

$npx convez-ai Dubibubii/claude-usage-widget

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Monitorar limite semanal e sessão de 5h em tempo real
  • Acompanhar uso do Agent SDK e créditos disponíveis
  • Analisar padrões de consumo de tokens com histórico de 14 dias
  • Acessar skills práticas para reduzir uso de tokens

Veredito

Ferramenta bem construída que resolve um problema real: visibilidade de uso Claude sem chamadas à API proibida. A implementação é rigorosa (capturas de statusline, reconstrução de histórico local, testes pixel-perfeitos), mas requer calibração contra faturamento real; a versão paga por temas é honesta, sem dark patterns.

README

A tradução para o português deste README ainda está na fila.

Claude Usage Widget

An always-on-top desktop widget that shows your Claude usage limits at a glance — the 5-hour session window, the weekly cap, and Agent SDK credits — snapped to a corner of your screen. Click to expand into a small dashboard.

Tracking is free, forever. The only paid feature ($5/mo) is cosmetic customization: alternate collapsed widget styles and skins.

Design: Glass Instrument — see design/ handoff bundle (spec board + README are the source of truth for pixels and timings; deliberate deviations are commented at the point of change).

Status

  • ✅ Native menu-bar app (macOS): accessory process (no Dock icon), tray donut with live %, frameless transparent pill above fullscreen apps, deterministic drag with corner snapping (in-app live self-test: CUW_SELFTEST=1, 9 scenarios, pixel-exact assertions), content-fit window, panel-only desktop blur. Windows build via CI (untested by hand).
  • ✅ Live data, no demo values. Two ToS-compliant local sources:
    1. Session + weekly %: a statusline shim (bundled: the app binary itself runs as --statusline; dev: scripts/claude-statusline-shim.sh) captures Claude Code's own rate_limits.{five_hour,seven_day} payload. Registered via statusLine in ~/.claude/settings.json. Never calls the banned /api/oauth/usage endpoint, never touches the OAuth token. Percentages refresh on terminal Claude turns only (the desktop app doesn't run statuslines); between turns the last value is shown as an honest floor until its window resets, labeled with its sync age.
    2. Local history / SDK credits / by-app / all-time tokens: reconstructed from ~/.claude/projects/**/*.jsonl (dedup by requestId, entrypoint classifies the SDK pool, cost derived from vendored pricing).
  • ✅ Pill model: every meter you keep (✕ in Edit toggles) rings in the collapsed pill; cards always stay in the expanded grid. Default = the 5-hr session ring.
  • ✅ History: 14-day chart of real daily tokens from your transcripts (populated on first launch) with observed weekly-peak overlay, plus Agent-SDK-$ by app.
  • ✅ Ambient insights (src/data/insights.ts, pure + unit-tested): cache-read share, dominant-project flag, weekly climb rate on the card, SDK pool pace alert ("~empty Jun 19 at this pace"), billing-cycle day in the footer, and a one-click weekly digest built to be handed to Claude.
  • ✅ Skills tab: starter pack of token-saving skills (context diet, cache hygiene, model routing, compact discipline, plan-first) exported as .md for ~/.claude/skills/ — the actionable half of "reduce my usage".
  • ✅ Exportable report: ~/Documents/claude-usage.md is a compact current-month report (daily token mix, cache-read share, model/project splits, pool usage, limit peaks) regenerated in place — built to be sent to Claude for token-saving analysis. Auto-resets monthly.
  • ✅ Licensing: Stripe checkout + Cloudflare Worker email validation (workers/license), 3-day revalidation, 14-day offline grace. Paid theme content lives in a private repo (src/paid/, gitignored); public builds fall back to Glass + Ring gracefully.
  • ⚠️ Calibration: the plan→SDK-pool mapping ($10/$50/$200) and the $-figures are API-equivalent estimates of Anthropic's metering, labeled "est." in the UI — validate against real billing.

Run

bun install
bun run dev        # browser dev mode at http://127.0.0.1:4733 (demo backdrop)
bun run tauri:dev  # desktop widget shell (needs Rust toolchain)
bun run tauri build  # release .app/.dmg (full build required to re-embed dist/)

Setup: live limit data (one step)

The 5-hr/weekly percentages come from Claude Code's own statusline. Register the bundled capture shim once:

./scripts/install.sh   # merges statusLine into ~/.claude/settings.json (backs it up,
                       # chains any statusline you already have), then restart Claude Code

After your next Claude turn in a terminal, the widget shows your real percentages. Until then it shows honest "waiting for first Claude turn" states — never demo numbers. (SDK credits, History, and the report work immediately: they read your local transcripts.)

Building from source vs official downloads

A source build includes the free Glass theme and Ring style — fully functional tracking, forever. The paid theme pack (5 skins + 6 collapsed styles, $5/mo) ships only in official release downloads and unlocks via license; the Theme tab in a source build shows previews. See docs/monetization.md.

Layout

src/
  state/      WidgetState model + reducer + persistence
  theme/      Design tokens (Dark/Light Glass) + auto OS theming + paid-pack loader
  data/       statusline capture parsing, transcript scanner, pricing, formatting
  widget/     Pill, Panel, 5 tabs, SetupCard, drag/snap + in-app self-test
  export/     Monthly usage report generator (~/Documents/claude-usage.md)
  platform/   Tauri window/drag/fs bridges (pure geometry unit-tested)
src-tauri/    Tauri v2 shell: tray, window level, native commands, --statusline mode
workers/      license validation (Cloudflare Worker + Stripe)

Privacy

Everything stays on this device. The usage report (claude-usage.md) is written locally and exported only by explicit user action. License activation sends only your checkout email to the validation endpoint.

Ver o original no GitHub

Bom para

Relacionados