$LLM Cost Barfor macOS
SPEND AND LIMITS · NEVER MORE THAN MINUTES STALE
$$0.11Mon 09:41

Your LLM spend and limits, live in your menu bar.

Track today's API burn across OpenRouter, Anthropic, and OpenAI — and your Claude Pro/Max and ChatGPT session/weekly limits with reset countdowns. Live charts, credit balances, per-API-key breakdowns. Free, open source, no backend.

Download for macOSView on GitHub
Free · Open source (MIT) · macOS 14+ · Signed & notarizedTotal downloads
LLM Cost Bar dropdown under the macOS menu bar: Today $0.11, MTD $73.84, API Spend and Subscriptions tabs, with Anthropic, OpenAI and OpenRouter rows
THE PROBLEM

Two ways to pay. Same blind spot.

Pay per token?

API keys at three vendors, agents running overnight, a credit balance draining quietly. The dashboards are three logins away, and none shows today — when the invoice arrives, answering what burned the money is a spreadsheet job.

See API spend →

Pay monthly?

On Claude Max or ChatGPT Plus, the question is worse: how much of my week is left? Right now that's answerable only by typing /usage into a terminal.

See your limits →
THE ANSWER — EITHER WAY

A menu bar number, never more than a few minutes stale — plus a dropdown that answers "which vendor, which key, which day" and "how close am I to the cap, and when does it reset" in two clicks.

Menu bar close-up: the LLM Cost Bar icon showing $0.11
WHAT IT DOES

Everything between "am I burning money?" and the answer.

A background daemon does the collecting; the menu bar shows the result. Nothing else to run, nothing to host.

Today, live

Not yesterday's report. A background daemon polls vendor APIs on your schedule, showing today's spend as it happens. For vendors with lagging activity feeds, a live delta keeps the number current.

Every vendor, one panel

One panel holds OpenRouter, Anthropic, and OpenAI cards, each with month-to-date, credits progress, and a 30-day bar chart. Hover any bar for that day's exact spend.

Which key is burning?

Per-key spend for every vendor: real dollars from OpenRouter and OpenAI, token-weighted estimates for Anthropic. Every key row shows three columns — today, month-to-date, trailing 30 days — aligned under the vendor's own totals.

Subscription limits, not just dollars

On Claude Max or ChatGPT Plus? A second tab shows every rolling limit window — Claude's 5-hour session and weekly caps (per-model too), Codex's weekly window — as color-coded bars with exact reset times.

Zero-setup detection

No keys to paste. If Claude Code or Codex CLI is signed in on your Mac, the app finds it and reuses that sign-in read-only — one "Always Allow" Keychain click for Claude, nothing for Codex. It never refreshes or touches the CLI's tokens.

Know before you hit the wall

A macOS notification fires when any window crosses your threshold (default 80%, configurable), re-arming after each reset. A 7-day burn-rate sparkline shows whether the week's budget will last until Friday.

Private by design

No backend. No telemetry. No account. Keys stay in the macOS Keychain; data flows directly from vendor APIs to a local SQLite file. The repo is open — inspect it.

Survives everything

A launchd-supervised daemon keeps collecting through app crashes, restarts, and sleep. Every sync attempt appears in Diagnostics, so nothing fails silently.

Set up in a minute

Paste one management/admin key per vendor — live-tested before saving, with a pointed error when a vendor hands you the wrong kind. Subscriptions need no setup at all: they're auto-detected.

Download for macOSFree · one .dmg · nothing to sign up for
INSIDE THE DROPDOWN · API SPEND

Which vendor, which key, which day — two clicks.

Expand any vendor card and the whole story is right there — no dashboard login, no CSV export.

30-day chart per vendor. Hover any bar for that day's exact spend — "Jul 12 · $3.41".

Color-coded credits bar. "$7.36 left of $80.00 · 90% used" — you see it drain before it's gone.

Per-key spend, three columns. Today, month-to-date, trailing 30 days for every key — your overnight agent, your IDE, your side project — so the culprit is named.

Fully expanded dropdown: Anthropic, OpenAI and OpenRouter cards with 30-day bar charts, three-column per-key spend lists, and OpenRouter's credits bar at 90% used
Subscriptions tab: Claude Max card with 5-hour and 7-day limit bars and reset countdowns, above a Codex card with its weekly window
INSIDE THE DROPDOWN · SUBSCRIPTIONS

How close am I to the cap — and when does it reset?

On Claude Max or ChatGPT Plus, spend isn't the question — the window is. The second tab shows every rolling limit as a color-coded bar with its exact reset time.

Every window, one glance. Claude's 5-hour session and 7-day caps (per-model too), Codex's weekly window — "61% — resets Thu 09:00 · in 2d 3h".

Zero-setup detection. If Claude Code or Codex CLI is signed in on this Mac, limits appear automatically — the sign-in is reused read-only, never modified or refreshed.

Know before you hit the wall. A notification fires at your threshold (default 80%), re-arming after each reset — and a 7-day sparkline shows whether the budget lasts until Friday.

PRIVACY

Your billing data never leaves your Mac.

There's no server because there doesn't need to be one. Your Mac calls the vendor APIs with your keys and keeps the results locally. That's the whole architecture.

Vendor APIsOpenRouter · Anthropic · OpenAI
your admin keysdirect connection
Your MacKeychain · local SQLite

No middle-man server. No telemetry. No account. Subscription limits reuse your CLI sign-ins read-only — nothing is modified, nothing sent anywhere new. See PRIVACY.md in the repo.

Keys stay in the Keychain

API keys are stored only in the macOS Keychain — never in config files, never uploaded anywhere.

Data lives in local SQLite

Spend history is a SQLite file on your disk. Query it yourself, back it up, or delete it — it's yours.

Open source, MIT

Don't take the privacy story on faith — the entire codebase is on GitHub. Read exactly what it does.

Settings window, General tab: what the menu bar shows, refresh interval, which tab the popover opens to, subscription alert threshold, keep-app-running and launch-at-login toggles
SETUP

One key per vendor — or no keys at all.

01

Download the .dmg, drag to Applications, open.

02

Tracking API spend? Paste one read-billing key per vendor. Setup links to the exact page where each key lives and live-tests it before saving — with a pointed error if a vendor hands you the wrong kind.

03

On subscriptions? Do nothing — if Claude Code or Codex CLI is signed in on this Mac, your limits are auto-detected.

04

Make it yours in Settings: what the menu bar shows, refresh cadence, which tab the popover opens to, the subscription alert threshold, launch at login.

OpenRouter · management keyAnthropic · Admin API keyOpenAI · Admin API keyClaude Pro/Max · auto-detectedChatGPT Plus/Pro · auto-detected
FAQ

Good questions.

Yes. MIT-licensed, no paid tier, no account. You bring your own vendor API keys — or none at all if you're only tracking subscriptions.
Read-billing keys, not your inference keys: an OpenRouter management key, an Anthropic Admin API key (team orgs — individual Console accounts can convert to team in one click), an OpenAI Admin API key (org owner). The in-app setup links to the exact page for each and tests the key before saving.
Yes — that's exactly what the Subscriptions tab is for. Skip the API-key setup entirely; if Claude Code or Codex CLI is signed in on your Mac, your limits appear automatically.
It reuses the sign-in Claude Code already stores in your macOS Keychain — read-only, never modified or refreshed, never sent anywhere except Anthropic's own API (the same endpoint Claude Code's /usage command calls). macOS asks once; click "Always Allow". If you'd rather not, one toggle turns the whole source off. Honest caveat: the limit endpoints are unofficial (the same ones the vendors' own CLIs use) — if a vendor changes them, cards show last-known data marked stale until an app update; spend tracking is unaffected.
Claude Pro/Max and ChatGPT meter usage in rolling windows (Claude: 5-hour session + 7-day; ChatGPT/Codex: weekly). Each bar is one window: percent used, plus exactly when it resets.
Nowhere. There's no server. Your Mac talks to the vendor APIs directly and stores results in a local SQLite database. Keys are Keychain-only. See PRIVACY.md in the repo.
The background daemon is a separate binary from the app, and macOS grants Keychain access per item — so you approve each newly added key once. Click "Always Allow" and you won't be asked again for that key.
Waiting on Google — AI Studio has spend dashboards but no API to read them. The moment one exists, it's a provider away.
The sandbox is a poor fit for a launchd daemon reading org billing APIs. It's distributed the way Alfred and Raycast are: a notarized, Developer ID-signed DMG straight from GitHub Releases.
$$0.11

Know what today cost before the invoice does.

One .dmg. One key per vendor for API spend — no keys at all for subscription limits. Today's burn and this week's window, a glance away — forever, for free.

Download for macOS
Free · Open source (MIT) · macOS 14+ · Signed & notarized