Tokenrunway

Features

• Local stats from your agents' logs: Claude Code, Codex, opencode and others. Spend per day, week, month and lifetime; your most-used models and projects; and what the work was (building, discovery, bug fixing…). Compare harnesses side by side.
• Local models too: LM Studio's requests, models and tokens, read from its server logs and shown as free.
• Catch runaway sessions: a warning when a session spends 3× your usual pace, so an agent loop doesn't eat your window.
• Switch to save: when a window is tight, it says where there's room (another model, or another seat).
• Spend goals: daily and monthly goals with rings on the Now tab and alerts at 80% and 100%.
• Use it or lose it: an optional reminder when a weekly or monthly allowance resets in the next 12 hours with 30% or more left, on your desktop and your Rattleboard.
• Insights: cost per project and git branch, cache-efficiency tips, a Monday recap of last week, and what your plan is worth at API prices ("your $200 plan did $874 of work, 4.4×").
• A full dashboard with daily spend by model, harness or category, a burn-down for any window, every session, upcoming resets and a contribution-style heatmap. Click a session to see its steps: what each cost, the tools it ran, the expensive steps and the cache misses. Click a model or project in Stats to open it there.
• Menubar your way: runway, % left, or today's spend, for the most urgent window or one you pin. A copy button puts a usage summary on the clipboard.
• tokenrunway on the command line: status, today, and a one-line statusline for your shell prompt or Claude Code.
• Sign in where there's no CLI: Cursor and Grok can be signed in to inside Tokenrunway. Only the session cookie is kept, in your keychain.
• Rattleboard: alerts and an "AI runway" card on a split-flap board.
• Current prices: the price list updates daily from GitHub, so new models are priced without an app update (the built-in prices stay as the fallback).
• Updates itself with release notes, and installs with Homebrew.
• Private: no account, no server, no telemetry. Your logins stay in your OS keychain, and your stats stay on your computer.

Screenshots

Screenshots use sample data.

Install

Platform	How
macOS 11+ (Apple Silicon and Intel)	brew install --cask slastrina/tap/tokenrunway, or the .dmg from Releases
Windows 10/11	the -setup.exe or .msi from Releases
Linux x86_64 / arm64	the .deb or .AppImage from Releases (needs libayatana-appindicator; on GNOME the AppIndicator extension)

Tokenrunway checks GitHub for new versions once a day, shows what's new, and updates in place when you say so. Homebrew installs update with brew upgrade --cask tokenrunway.

Command line

The app doubles as a tokenrunway command. Homebrew puts it on your PATH; the .deb installs it in /usr/bin; on Windows it's tokenrunway.exe in the install folder, and the AppImage takes the same arguments.

$ tokenrunway status          # every account's windows, resets and forecasts, and today's spend
$ tokenrunway statusline      # ▲ 1h 26m · Claude 5-hour 78% · $12.40 today
$ tokenrunway today           # today's spend by harness, model and project
$ tokenrunway status --json   # the same for scripts (also: today --json)
$ tokenrunway guard           # exit 0 if every window is under 90%, 1 with the reason if not

Homebrew also adds two shorter names for the same command, runway and trw (not tr, which is the Unix tool): runway status, trw today, runway guard.

Guard a script or agent loop: guard prints nothing when it's safe, so it chains:

tokenrunway guard --max 80 && claude -p "fix the flaky tests"
tokenrunway guard --account work --window 5h --runout || echo "out of runway, try later"

--max sets the limit (% used, default 90), --account and --window narrow it (part of the name, or 5h, weekly), --runout also stops when a window is forecast to run out before it resets, -v says what it checked and --json gives the details. Exit codes: 0 go, 1 over the limit, 2 bad options, 3 can't tell (the app isn't running, or nothing matches); --allow-unknown turns 3 into 0.

Allowances come from the running app (it shares them in its data folder), so start Tokenrunway first; spend comes straight from your agents' logs. Set NO_COLOR=1 to turn colours off.

Claude Code status line: add this to ~/.claude/settings.json:

{
  "statusLine": { "type": "command", "command": "tokenrunway statusline" }
}

Rattleboard

Tokenrunway can show its alerts, and an "AI runway" card, on a Rattleboard split-flap board. In Settings → Rattleboard:

1. Rattleboard on this computer: click Connect next to Found on this computer. Turn on Rattleboard's Settings → API server first. Alerts go through the Rattleboard app, with its own key.
2. A board on your network: in that Rattleboard, Settings → API server → Copy link (turn on Other devices), paste the rattleboard://pair?… link and click Connect. Or open or enter address and key.
3. Test shows "TOKENRUNWAY CONNECTED" for 15 seconds. Pick which alerts go to the board, and turn on Show an AI runway card in the rotation. The card needs the board's key: on this computer, paste its pair link under Add the key for the card.

The key is kept in your OS keychain. The board only gets window names, percentages and reset times. It never gets prompts or costs. Disconnect removes the card from the board and the key from the keychain.

Build from source

npm ci
npm run dev        # run it
npm run build      # installers in target/release/bundle
cargo test --workspace

Linux needs libwebkit2gtk-4.1-dev libayatana-appindicator3-dev librsvg2-dev libdbus-1-dev libxdo-dev.

Layout: crates/tokenrunway-core is the engine (providers, log ingest, pricing, forecasts, update checks) with no Tauri dependency; src-tauri is the desktop app; src is its UI (plain HTML/JS).

Releasing

1. scripts/bump.sh 0.2.0
2. Add ## [0.2.0] - YYYY-MM-DD to CHANGELOG.md. That section becomes the release notes and the in-app "What's new".
3. Commit, then git tag v0.2.0 && git push --tags. GitHub Actions builds and signs every platform, publishes the release and latest.json for the updater, and bumps the Homebrew cask.

Unofficial endpoints

Some providers have no public usage API for subscriptions. For those, Tokenrunway reads what their own CLIs already keep on your computer, or calls the endpoints their own apps call, for your own account only. Those can change without notice. When one does, Tokenrunway keeps the last good reading and says the source needs an update. The Gemini, Grok and xAI readers were built from the providers' observed calls without a test account, so treat them as early.

Licence

MIT. Not affiliated with Anthropic, OpenAI, Amazon, xAI, Google, GitHub or Cursor.