TokenEater

Requires a Claude Pro, Max, or Team plan, or the Codex CLI signed in with ChatGPT. Either one is enough. Claude's free plan does not expose usage data, and Codex API-key accounts have no usage windows to track.

What you get

A native menu bar app, desktop widgets, and a floating overlay that track your Claude and OpenAI Codex usage in real time, side by side or one at a time.

• Menu bar. Live percentages with color-coded thresholds, and a popover dashboard you compose element by element (rings, chips, arcs, pacing bars at full, half, or third width), start from built-in templates, and save as your own.
• Dashboard. A four-space window (Monitoring / History / Studio / Settings) with flippable tiles, 7-day sparklines, peak day, and a pacing-vs-equilibrium graph. Its blocks reorder and hide from Studio, like the popover and the menu bar.
• Two providers, one app. Claude and OpenAI Codex are equals: the menu bar, the popover, the dashboard, widgets, notifications and History cover both. An All / Claude / Codex mode scopes every surface at once, All puts the two side by side, and each mode keeps its own layouts. Track only one and you never see a trace of the other.
• History. Combined Claude Code and Codex tokens from local session logs: a stacked chart by model, project ranking, session counts, and cache hit rate, filterable by Claude family or Codex model across 24h to 90d ranges. The History widget shows the combined daily totals.
• Widgets. Native WidgetKit gauges, progress bars, and pacing, refreshed reactively, with a Codex Usage widget alongside the Claude ones.
• Agent Watchers. A floating overlay of your live Claude Code sessions, terminals and VSCode-family extensions alike. Click a session to jump to its terminal or editor (Terminal, iTerm2, tmux, Kitty, WezTerm), right-click for quick actions.
• Smart Color. Blends how much you have used with how fast you are burning, so the color warns you before the number does. Three temperaments set how cautious it is.
• Smart pacing. Are you burning through tokens or cruising? Four zones: chill, on track, warning, hot.
• Themes. Four presets plus full custom colors, a glow or flat look, and configurable warning thresholds.
• Notifications. Per-event toggles for each provider: escalation, recovery, pacing, scheduled reset reminders, extra credits, token expiry. Every alert names the provider it is about.

Everything in detail on the website.

Install

Download DMG (recommended)

Download TokenEater.dmg

Open the DMG, drag TokenEater to Applications, and launch it. The DMG is signed with a Developer ID and notarized by Apple, so Gatekeeper lets it run on first launch without any extra steps.

Homebrew

brew tap AThevon/tokeneater
brew trust AThevon/tokeneater
brew install --cask tokeneater

brew trust is required on Homebrew 6.0+, which no longer loads a third-party tap until you trust it.

First setup

Prerequisites: at least one of Claude Code signed in on a Pro, Max, or Team plan (claude then /login), or the Codex CLI signed in with your ChatGPT account (codex login).

1. Open TokenEater: a guided setup detects the providers on your Mac and walks you through connecting the ones you use
2. Right-click on the desktop > Edit Widgets > search "TokenEater"

Update

TokenEater checks for updates automatically. When a new version is available, a modal lets you download and install it in-app; macOS will ask for your admin password to replace the app in /Applications.

If you installed via Homebrew: brew update && brew upgrade --cask tokeneater

Uninstall

Delete TokenEater.app from Applications, then optionally clean up shared data:

rm -rf /Applications/TokenEater.app
rm -rf ~/Library/Application\ Support/com.tokeneater.shared

If you installed via Homebrew: brew uninstall --cask tokeneater. For a complete wipe, caches and widget state included, use the clean reset in the troubleshooting guide.

Build it yourself

git clone https://github.com/AThevon/TokenEater.git
cd TokenEater
./build.sh

The script checks Xcode, installs XcodeGen if needed, and assembles the app. Local builds are not notarized, so Gatekeeper blocks the first launch (right-click > Open, or System Settings > Privacy & Security > Open Anyway). The step-by-step walkthrough is in SETUP.md.

Privacy: read-only usage calls

TokenEater reads the OAuth access token Claude Code already keeps in your macOS Keychain, the same token Claude Code itself uses. At first launch, macOS asks you to allow that access: click Always Allow once. The prompt is standard macOS behavior for any app reading a keychain item it did not create, and since the read goes through Apple's own security tool, whose signature never changes, the prompt does not come back on updates.

Everything the app does with the token:

• GET api.anthropic.com/api/oauth/usage, your current usage stats
• GET api.anthropic.com/api/oauth/profile, your plan info

When Codex tracking is enabled, the app also reads the Codex CLI access token from ~/.codex/auth.json (or CODEX_HOME) and makes one additional authenticated call:

• GET chatgpt.com/backend-api/wham/usage, your Codex usage windows and credits

These usage and profile calls are read-only. TokenEater does not send messages, read conversations, or modify either account. Each access token is sent only to its own provider. TokenEater never writes to either provider's credentials and never refreshes their tokens: if one expires, open Claude Code or Codex (or run codex login) and it picks the new one up. API-key and keyring-only Codex logins are not supported in this version.

Widgets read local JSON files with no network or Keychain access. The Codex widget reads codex.json for usage and shared.json for theme preferences; its cache contains no token, email, or account ID. History reads Claude Code and Codex local session logs without uploading them; Agent Watchers reads Claude Code sessions.

Anthropic does not offer a third-party OAuth flow or scoped tokens yet, so reading the existing token is the only way an app like this can exist. If scoped tokens become available, TokenEater will adopt them immediately. The relevant code is short and auditable: keychain access in SecurityCLIReader.swift and TokenProvider.swift, the Claude calls in APIClient.swift, and Codex access in CodexAuthReader.swift and CodexAPIClient.swift.

Codex setup and widgets

Sign in to the Codex CLI with your ChatGPT account. TokenEater enables Codex tracking automatically on the first launch that detects that login. If you sign in later, turn it on under Settings > Providers, where each provider has its own tracking switch and the coverage matrix shows what each feature does on each one. Agent Watchers stay a Claude Code feature: they read Claude Code's own session processes, which Codex has no equivalent of. The onboarding wizard sets up either provider or both, and you can finish it with OpenAI alone.

History includes local sessions from both providers, independently of quota authentication. It reads Codex sessions/ and archived_sessions/ under CODEX_HOME (default ~/.codex). All sums uncached input, cache writes, and output tokens for both providers. Only cache reads count as reused tokens in the cache breakdown. Guardian / auto-review s