Pulse shows how much Claude Code, Codex, Cursor and 17 other AI coding subscriptions you have left, on the edge of your Mac screen

The developer qunqin24 published Pulse, an Apache-2.0 macOS menu-bar app that reads what is left in your AI coding subscriptions and shows each one as a coloured ring docked to the edge of the screen. The repository was created on 2026-08-30 and reached 407 stars and 47 forks in three weeks. It runs on macOS 14 Sonoma or newer, Apple Silicon and Intel, and is written in Swift 6.

What it reads and how

The rail covers twenty providers today: Claude Code, Codex, Kiro, Antigravity, Cursor, GitHub Copilot, Grok, Grok Bot, OpenCode Go, Kimi Code, Ollama Cloud, z.ai, Zhipu, MiniMax in the international and mainland versions, Volcengine, Command Code, DeepSeek, Devin and the Xiaomi Coding Plan. Each ring is what that service's own API says is left in the current window, read through the client route you already signed in with. The project's own words: "Pulse runs on your Mac under your own provider logins", with nothing reported back to a Pulse server. A ring turns from green to amber to red as the quota empties, with a countdown mode that flips between spent (75% used) and remaining (25% left).

Hover a ring and Pulse expands a card with every quota pool the provider reports, the reset countdown and the current window. An optional burn-rate forecast, off by default, extrapolates the current pace and shows when the quota is projected to run out. Pin the quota that matters most, or let Pulse pin the one closest to zero on its own. Notifications are all opt-in and each fires once: a limit crossing 75, 80, 90 or 95 percent, the provider marking a limit spent, a window reopening, or, for services that sell prepaid credit, the balance falling below a threshold you set.

The token-spend pane, off by default

A separate feature, off until you switch it on, scans local logs and databases from a catalogue the README puts at 54 client sources. That list includes Gemini CLI, Cline, Roo Code, OpenClaw and GitHub Copilot; Cursor, Trae and other export-only sources need a prior export or capture. The pane opens on the last seven days by default, prices tokens against the published API rates rather than the subscription rate, and labels an unknown price as unavailable rather than guessing. Individual model views show input, output and cache counts, contributions by agent, and daily and hourly charts where the source has the timing to draw them. Coverage varies by source and the project says so.

The command that other panels can read

Pulse --json prints the last cached readings for every account: plan, every limit, reset times and how old the figures are. The project suggests it for tmux, sketchybar, Raycast or a shell prompt. Because it reads the cache rather than triggering fetches, polling it costs nothing on the provider side. Pulse also exports a Raycast extension and ready-made scripts for tmux, sketchybar and the shell from the Settings pane. Five interface languages, English, Simplified Chinese, Traditional Chinese, Japanese and Korean, use language-aware number units (K/M/B, 万/亿, 萬/億, 万/億, 만/억).

Two or three paid AI coding subscriptions is now normal for a working developer, and hitting a limit halfway through a task is the common cost. Most people either guess from the last error, or open each provider's dashboard in a separate browser tab. A floating rail changes that: the choice of which tool to reach for on the next turn is made against a real number, and running out of one becomes a decision rather than an accident.

Source

Primary source: qunqin24/Pulse on GitHub.