StatStrip

Live CPU, RAM, disk, GPU — and your Claude Code + OpenAI Codex usage limits — inside the Windows taskbar.

No extra window, no wasted screen space: the numbers sit in the empty part of the taskbar itself.

^{That's the real thing, in a real taskbar — not a mockup.}

If you use Claude Code or Codex all day, the question you actually care about isn't CPU — it's how much of my 5-hour block have I burned? StatStrip keeps that answer on screen, next to your hardware stats, without spending any of the quota it's reporting.

• Real plan limits, not estimates. Percent-of-your-actual-limit, read from the same sources the tools use themselves.
• Costs you nothing to watch. The Codex gauges run no model — zero tokens. The Claude gauges reuse your existing login.
• Never lies to you. A number it can't verify is marked with its age or hidden outright. Stale figures that look live are treated as bugs, not features.
• Nothing leaves your PC. No API keys, no accounts, no telemetry, no server.

Install (2 minutes, no technical knowledge needed)

1. Download: click the green Code button at the top of this page → Download ZIP, then right-click the file → Extract All.
2. Run: open the extracted folder and double-click install.bat.

That's it. The installer finds Python (and downloads it automatically if it's missing — this can take a few minutes the first time), installs StatStrip, and starts it. The stats appear in your taskbar right away and come back every time you log in. To remove it, double-click uninstall.bat in the same folder.

If the automatic Python download fails (older Windows without winget), install Python once from python.org/downloads — tick "Add python.exe to PATH" — and run install.bat again.

What the gauges mean

Codex names its own windows, and StatStrip shows whatever your plan reports rather than assuming. On a Plus plan that's a single weekly window (CODEX 7d 14%); other plans report more, including a 5-hour one.

Privacy — no API keys, nothing leaves your PC

• No API keys, no accounts, no signup. The Claude gauges reuse the Claude Code login already on your PC — the same request Claude Code makes when you run /usage. No Claude Code? They show "login required"; set STATSTRIP_CLAUDE=off to hide them.
• The Codex gauges don't spend any of your quota. They ask your local codex CLI for your usage (account/rateLimits/read, the same call behind its /status), which runs no model and costs zero tokens.
• Your stats stay on your machine. Hardware numbers are read locally and served only on 127.0.0.1. Nothing is uploaded, logged, or shared — there is no server, no telemetry, no third party.

Honest numbers

Most of the engineering here is about not showing a plausible wrong number. A usage gauge that quietly drifts is worse than no gauge, because you'll trust it.

• Readings carry the instant they were taken, and the display ages them itself. A collector that wedges shows a reading growing visibly older ((3h ago)), never one frozen at "just now".
• A window that has since reset shows reset, not a number known to be stale.
• "Login required" is reserved for an actual missing login. A rate limit, a network blip, or an endpoint hiccup says usage unavailable instead — sending you off to re-authenticate for a problem that isn't about auth is a bug.
• StatStrip backs off instead of making things worse. Claude's usage endpoint rate-limits; polling straight through a 429 keeps the limit alive, so failures back off exponentially and let it clear.
• Anything outside a source's expected contract is refused, not guessed at — an unparseable timestamp, a percentage outside 0–100, a renamed field.

Configuration

All optional, set as environment variables before launching.

Architecture

Two independent layers, so the data is reusable outside this app — point any script, dashboard, or website at the same local endpoint.

┌──────────────────┐      writes/serves       ┌─────────────────┐
│    collector     │ ───────────────────────▶ │  stats.json      │
│  (psutil/pynvml, │                          │  127.0.0.1:5757  │
│   Claude + Codex │◀───────── read only ─────┤      /stats      │
│      usage)      │                          └────────┬─────────┘
└──────────────────┘                                   │
                                             ┌─────────▼──────────┐
                                             │      display        │
                                             │ (taskbar bar, or    │
                                             │  your own consumer) │
                                             └─────────────────────┘

• collector.py — extraction only. Polls hardware, Claude and Codex usage, then writes a merged snapshot to a JSON file and serves it on http://127.0.0.1:5757/stats.
• display.py — consumption only. Renders the bar embedded inside the taskbar (TrafficMonitor-style, transparent, left of the tray icons). It has zero knowledge of collection internals — it's just one consumer of the feed.
• claude_oauth.py — primary Claude source: the same endpoint Claude Code's /usage uses, authenticated with the token it already stores locally.
• claude_local.py — opt-in estimate mode (STATSTRIP_CLAUDE=estimate): shells out to ccusage over your local logs and estimates against your own historical maximum (can exceed 100% — it's an estimate, not a plan limit).
• codex_appserver.py — primary Codex source: asks codex app-server for account/rateLimits/read, the call the Codex TUI makes to refresh its own usage. Runs no model, so it's free to poll — and it reports your whole account, including Codex you ran on another machine.
• codex_local.py — fallback Codex source (STATSTRIP_CODEX=log forces it): reads the rate-limit snapshot Codex writes into its session logs. Fully passive, but only sees turns run on this machine — measured 6 points behind the live figure — so its readings carry an age marker.

Anything else — a browser tab, another script, an Electron app — can hit the same /stats endpoint or read the JSON directly. (Browser pages need STATSTRIP_CORS; it's off by default.)

Development

Requires Python 3.9+ on Windows. NVIDIA GPU optional — it degrades gracefully.

git clone https://github.com/wassupjay/statstrip.git
cd statstrip
pip install .
statstrip-collector
statstrip-display

Run the tests (no dependencies beyond the package itself):

python -m unittest discover -s tests

Contributions welcome — issues and PRs both. The one house rule is the one above: a gauge must never show a number it can't stand behind. If a source fails, is stale, or returns something outside its contract, say so or say nothing.

License

MIT — see LICENSE.