claude-usage

README

claude-usage

Know how much of your Claude.ai plan you've burned — from your terminal, shell prompt, or menu bar.

claude-usage is a tiny cross-platform CLI that reads the Plan Usage numbers from your Claude.ai account (Current session, Weekly all-models, Sonnet, Design) and surfaces them where you actually work: a one-shot status command, a live TUI, a shell-prompt segment, or a system tray icon.

[demo gif goes here — record with vhs after Phase 1]

Numbers are battery-style: the percentage and bar represent how much capacity is left in each bucket (100% = untouched, 0% = at the plan ceiling). Threshold colors are still decided from raw usage so warn/alert fire at the same place they always did.

$ claude-usage status
Session   █████████░░░░░░░░░░░  49%  (resets in 1h 5m)
Weekly    █████████████████░░░  87%  (resets Sun 3:00 PM)
Sonnet    ████████████████████ 100%
Design    ████████████████████ 100%

---

Install

macOS

brew install tonydisco/tap/claude-usage

If brew fails with "Your Xcode is too outdated"

claude-usage is a pre-built bottle and does not need Xcode — just the Command Line Tools. Point xcode-select at the CLT and retry:

sudo xcode-select -s /Library/Developer/CommandLineTools
brew install tonydisco/tap/claude-usage

If CLT is not installed: sudo xcode-select --install.

Windows

# PowerShell one-liner (recommended)
irm https://raw.githubusercontent.com/tonydisco/claude-usage/main/install.ps1 | iex

# Or via Scoop
scoop bucket add tonydisco https://github.com/tonydisco/scoop-bucket
scoop install claude-usage

The installer drops claude-usage.exe into %LOCALAPPDATA%\claude-usage\bin and adds it to your user PATH. Open a new terminal afterwards.

Linux

curl -fsSL https://raw.githubusercontent.com/tonydisco/claude-usage/main/install.sh | bash

Any platform (from source)

go install github.com/tonydisco/claude-usage/cmd/claude-usage@latest

Binaries are signed and published from GitHub Actions via goreleaser. Source: .github/workflows/release.yml.

Tray / daemon. macOS release binaries (Homebrew Cask) are CGO_ENABLED=1 builds linking AppKit, so claude-usage tray and claude-usage daemon start work out of the box on macOS. Linux and Windows binaries remain CGO_ENABLED=0 for clean cross-compilation — to get the menu-bar/system-tray icon on Linux, build from source with a C toolchain available:

go install github.com/tonydisco/claude-usage/cmd/claude-usage@latest

Everything else (status, watch, prompt, login, …) works in every build.

Quick start

# 1. Authorize — scans your browsers for an existing claude.ai session,
#    or opens claude.ai and auto-detects once you sign in. No copy/paste.
claude-usage login
# (use `claude-usage login --manual` on headless boxes to paste manually)

# 2. Check status
claude-usage status

# 3. Embed in your zsh prompt
PROMPT='%~ $(claude-usage prompt) ❯ '
# → ~/code [51%/13%] ❯

# 4. (optional) Tray icon + background notifications
claude-usage daemon start

Features

• Cross-platform single binary — macOS (Intel + Apple Silicon), Windows, Linux. No runtime to install.
• Three frontends in one tool — CLI, live TUI (watch), shell-prompt segment (prompt), and tray icon (daemon).
• Tracks Claude.ai plan usage (Pro / Max), not just Claude Code token costs.
• Secure auth — session cookie is stored in macOS Keychain / Windows Credential Manager via zalando/go-keyring, never written to a plaintext config file.
• Configurable thresholds & notifications — warn at 80% / 95% by default, override in ~/.config/claude-usage/config.toml.
• oh-my-zsh plugin included — drop into $ZSH_CUSTOM/plugins/claude-usage and add claude-usage to your plugins=(...).

How is this different from existing tools?

There are several great Claude-related usage trackers already. They mostly target Claude Code's local JSONL files. claude-usage targets the Plan dashboard on claude.ai itself, which is the number that actually decides whether you can keep chatting today.

Tool	What it tracks	Stack	Platforms	UI surfaces
claude-usage (this)	claude.ai plan usage (session + weekly buckets)	Go, single binary	macOS · Windows · Linux	CLI · TUI · shell prompt · tray
ccusage	Claude Code token cost (local JSONL)	Node/TS, npx	Any with Node	CLI · statusline
Claude-Code-Usage-Monitor	Claude Code real-time, with ML predictions	Python	Any with Python	Terminal TUI
Claude-Usage-Tracker	claude.ai plan usage	Swift / SwiftUI	macOS only	Menu bar (native)
Claude-Usage-Extension	claude.ai plan usage	Browser extension	Browser only	In-page overlay

Pick claude-usage if you live in a terminal and want plan info in your shell prompt, or you need a cross-platform tray icon on Windows/Linux.

Pick ccusage if you mostly use Claude Code from the CLI and care about token cost per project.

Pick Claude-Usage-Tracker if you're macOS-only and want a polished native menu bar app.

Commands

Command	Purpose
claude-usage login	Paste session cookie, store it in OS keychain
claude-usage status	Print current usage once (colored when TTY)
claude-usage watch	Live TUI, refreshes every 60s, r to refresh manually
claude-usage prompt	Compact [51%/13%] for embedding in PS1
claude-usage daemon start|stop|status	Background process + tray icon
claude-usage config <key> <value>	Edit threshold / poll interval / theme
claude-usage logout	Remove credential from keychain
claude-usage completion zsh|bash|fish	Generate shell completion

Configuration

~/.config/claude-usage/config.toml:

poll_interval_seconds = 60
warn_threshold = 80   # show yellow when any bucket >= 80%
alert_threshold = 95  # show red + notify when any bucket >= 95%
notify = true
org_id = ""           # leave empty to auto-detect; set if you have multiple orgs

How it works (and the honest caveat)

Anthropic doesn't publish an API for Claude.ai plan usage. This tool calls the same internal endpoint that the Settings → Usage page in your browser hits, authenticated with the session cookie picked up by login.

claude-usage login reads the sessionKey cookie directly from your installed browsers (Firefox, Safari, Chrome, Edge, Brave, …) via browserutils/kooky. On macOS, reading Chrome cookies asks the OS Keychain for the encryption key — you'll see a one-time "claude-usage wants to access Keychain" prompt; pick Always Allow to silence it. If you'd rather not let the tool read browser storage, use claude-usage login --manual and paste the cookie yourself.

This means:

• Anthropic may change the endpoint shape at any time. When that happens the tool will likely break until I ship a patch. Watch Releases — patches usually land within 24–48h.
• Your session cookie is sensitive (it's a login). claude-usage stores it only in your OS keychain, and only the HTTPS request to claude.ai ever sees it.
• This is not affiliated with or endorsed by Anthropic.

The fetcher is isolated in internal/fetcher so endpoint patches are a single-file change.

Roadmap

• Phase 0 — Endpoint reconnaissance (see samples/usage-response.json)
• Phase 1 — Core CLI (login, status, logout)
• Phase 2 — TUI (watch) + shell prompt + oh-my-zsh plugin
• Phase 3 — Daemon + tray icon + threshold notifications
• Phase 4 — Homebrew tap + install script (Scoop bucket and Apple notarization pending)

Contributing

PRs welcome — especially for:

• Endpoint patches when claude.ai changes things (see internal/fetcher/)
• Linux tray icon polish (Wayland is the hard part)
• Translations of CLI output

Run the test suite with go test ./.... Mock data lives in samples/.

License

MIT — see LICENSE.