README.md

May 5, 2026 · View on GitHub

TrackNotch

TrackNotch

Real-time LLM usage tracking — right in your Mac's notch.


TrackNotch is a native macOS app that monitors your LLM usage across Claude, OpenAI, Cursor, Codex, and more — and surfaces it in the notch (or top of the menu bar on non-notched Macs). No proxies. No cookies. No telemetry.

Features

Local-firstReads usage directly from providers' own files and APIs. Nothing leaves your machine.
Multi-providerClaude Code, OpenAI API, Cursor, Codex, Anthropic API, Google Gemini — all in one pill.
Context arcVisual arc shows how full your active Claude session's context window is, live.
Budget trackingSet monthly budgets for OpenAI (admin key) and Anthropic (org admin key). See spend at a glance.
Rate-limit headersOAuth token support for real 5h/7d Claude rate-limit data from Anthropic's headers.
Notch-nativeSlides out of the notch with a springy open animation and clean ease-in close.
Menu bar fallbackWorks on non-notched Macs too — sits cleanly in the menu bar.
Keychain storageAPI keys stored in macOS Keychain. Never written to disk in plaintext.

Supported Providers

ProviderTracks
Claude CodeSession context usage, 5h/7d rate limits via OAuth
Anthropic APIMonthly org-wide spend — requires an Admin key (sk-ant-admin-…). Individual API keys not supported.
OpenAI APIMonthly API spend — requires an Admin key (sk-admin-…)
CursorSubscription fast-request usage
CodexSession usage

Install

  1. Download the latest TrackNotch-x.y.z.dmg from Releases.
  2. Open the DMG and drag TrackNotch.app into Applications.
  3. First launch: Because this build is unsigned, macOS Gatekeeper will block it.
    • In Finder → Applications → right-click TrackNotch.appOpenOpen.
    • This one-time step is enough. After that, launch normally.
    • If still blocked: xattr -cr /Applications/TrackNotch.app in Terminal, then retry.

Signed + notarized build is on the roadmap — no more right-click needed.

Setup

  1. Launch TrackNotch — the pill appears at the top of your screen.
  2. Click the pill → dropdown opens → Settings.
  3. Paste API keys for the providers you want to track. For Anthropic and OpenAI cost tracking, admin-level keys are required.
  4. For Claude Code rate-limit tracking, add your OAuth token (Settings → Claude Code → Rate-limit tracking).
  5. Usage refreshes automatically and the pill updates live.

Claude Code context arc

TrackNotch reads Claude Code's local JSONL session files to calculate context usage. No token or key needed.

To also get real 5h/7d rate-limit data, add an OAuth token:

# Install Claude Code if you haven't already
npm install -g @anthropic-ai/claude-code

# Generate an OAuth token
claude setup-token

Paste the token (starts with sk-ant-oat01-…) into Settings → Claude Code → Rate-limit tracking.

Requirements

  • macOS 13 (Ventura) or later
  • Apple Silicon or Intel Mac
  • API keys for providers you want to track (Claude Code and Cursor work via local monitoring without keys)

Privacy

All data stays local. Full details in PRIVACY.md.

  • API keys → macOS Keychain only
  • Local reads scoped to provider data dirs (~/.claude, etc.)
  • Network requests only to provider APIs you configure
  • Zero analytics, zero telemetry, zero third-party services

Why is the app unsandboxed?

TrackNotch reads provider data from locations like ~/.claude that are outside the App Sandbox. Shipping unsandboxed avoids per-launch folder-access prompts. The tradeoff is documented and intentional — see PRIVACY.md.

Building from Source

git clone https://github.com/manojacharix/tracknotch.git
cd tracknotch/TrackNotch
open TrackNotch.xcodeproj

Requires Xcode 15+ and macOS 13 SDK. Press ⌘R to run, ⌘U for the test suite.

# Build a DMG locally
./scripts/build-release.sh
# Output: build/TrackNotch-<version>.dmg

Roadmap

  • Apple Developer ID signing + notarization
  • Auto-update via Sparkle
  • First-launch onboarding flow
  • Light-mode support
  • Crash logs in ~/Library/Logs/TrackNotch/

Contributing

Bug reports and PRs welcome. Open an issue first for anything larger than a small fix.

License

MIT © 2026 Manoj Achari