Fire off a prompt, go do something else. When the animal stops running, your AI is done.
You know the loop. You type the prompt. You tab away. Ten seconds later you tab back — still going. Tab away. Tab back. Still going. You end up babysitting the thing you delegated so you wouldn't have to babysit it.
VibeCheck puts a tiny pixel animal on your screen that runs only while your coding assistant is actually working. Now you just glance. Still running? Still working. Stopped? Go look. That's the whole product.
git clone https://github.com/b2narae/VibeCheck.git &&cd VibeCheck
./scripts/make-app.sh && cp -R build/VibeCheck.app /Applications/ && open /Applications/VibeCheck.appThat's it. No config file, no API key, no sign-in. It finds your sessions by itself. Works with Claude Code, Gemini CLI, and Codex CLI.
🐎 One animal per terminal. Running three sessions at once? You get three animals — a horse, a turtle, a tiny dinosaur — each one bound to a specific terminal. Suddenly "which of my five tabs is still cooking?" is a question you answer by looking, not by clicking through tabs.
🧱 A wall means it wants something from you. When Claude stops to ask for permission, or wants an API key, or asks you a question, a brick wall drops in front of that animal and it halts, shoving against the wall until you come back. And the moment you answer in the terminal, the wall comes down — it tracks the session's actual transcript, not just the notification that raised it.
🖱️ Click an animal, see where it stands. Hover one and click — which
project it's in, what you asked, and the latest thing it said back. If it's
stopped at the wall, the panel shows the actual question it's asking or the
exact tool call it wants permission for (Bash — git push …). The wall itself
is clickable too.
🐾 The menu bar is a headcount. The status item shows one animal per running session — blocked ones lead with their wall, like 🧱🐕. Hover a session's row in the menu for the same details as clicking its runner.
📉 Height = how much Claude you have left. Claude's usage limit refreshes in 5-hour blocks. A fresh block puts your runner up near the top of the screen, and it drifts lower — and fades — as you burn through the window. When the window is fully spent, the runner stops as a tiny tombstone until a fresh block begins. Running along the bottom? Wrap it up, the reset is coming.
🔔 Sounds you choose. Pick any macOS system sound for "done" and for "needs you" — or silence. It previews as you pick.
💤 It gets out of the way. Clicks pass straight through the animals, so nothing blocks the app you're working in. No Dock icon. Turn the overlay off entirely and keep just the menu bar readout if you want.
Pick your favorite per assistant, or set it to 🎲 and let every terminal surprise you. Each one is a hand-drawn pixel sprite with a real four-frame run cycle — legs actually move.
| 🐎 Horse | 🦄 Unicorn | 🐫 Camel | 🐕 Dog |
| 🐈 Cat | 🐇 Rabbit | 🐢 Turtle | 🦖 T-Rex |
| 🐖 Pig | 🐄 Cow | 🦌 Deer | 🦘 Kangaroo |
| 🐆 Cheetah | 🐿️ Squirrel |
Yes, the turtle runs at the same speed as the cheetah. Life is unfair.
Out of the box, VibeCheck works everything out by watching your machine — no setup required. But if you want it to be precise, flip on "Claude Code 훅 연동" in the menu.
That lets Claude Code tell VibeCheck directly when a turn starts, when it ends, and when it's stuck waiting on you — instead of VibeCheck inferring it. Permission prompts raise the wall the instant they appear.
/Applications/VibeCheck.app/Contents/MacOS/VibeCheck --install-hooks # or use the menu
/Applications/VibeCheck.app/Contents/MacOS/VibeCheck --uninstall-hooksIt merges into ~/.claude/settings.json without touching your other settings or
hooks, and backs the file up first. Turning it off removes only what VibeCheck
added.
Will this slow my machine down? It's a single native Swift binary — no Electron, no runtime, no dependencies. It checks your process list a couple of times a second and draws some pixels.
Does it phone home? No. There is no network code in this app at all. It reads your local session files and that's the end of it.
Does the hook slow down Claude? Hooks block Claude while they run, so VibeCheck registers only turn-boundary events (not per-tool ones) and its hook finishes in about 10 ms.
I don't use Claude Code. Gemini CLI and Codex CLI runners work too. The usage-window height and the wall are Claude-specific for now.
Something looks wrong. Run ./build/VibeCheck --debug — it prints what it thinks
each session is doing, once per poll.
For the curious (click to expand)
VibeCheck polls the process list every 1.5 seconds and tracks each claude /
gemini / codex process by PID. Only interactive terminal sessions count:
resident helpers (daemons, pty hosts), one-shot invocations (claude auth status, claude -p …), and CLIs spawned from another session's own tools are
all skipped, and a new pid must survive two polls before it may appear — so a
dev server shelling out to gemini auth status every few seconds doesn't flood
your screen with phantom runners. Working directories come from libproc
(proc_pidinfo), so no subprocess is spawned per poll.
With hooks installed, SessionStart, UserPromptSubmit, Notification,
Stop and SessionEnd each run VibeCheck --hook, which records that session's
phase under ~/.vibecheck/sessions/. Only genuine "blocked on you" notifications
raise the wall — idle reminders and completion notices don't. Hook reports are
trusted only while the transcript agrees with them: interrupting a turn with Esc
fires no Stop hook, so a "working" report whose transcript has been quiet for
three minutes is distrusted, and approving a permission prompt fires no hook
either, so an "attention" report is dropped as soon as the transcript moves past
it. Hooks also carry the session's real id, which pins log lookups to the right
file when several sessions share one project.
The click panel's details come from the transcript too: the most recent
assistant text is the "latest response", and an unanswered tool_use at the end
of the log is the pending question (for AskUserQuestion, the question text
itself) or the tool call awaiting permission.
Without hooks, two signals are combined:
- CPU, summed across the session's child processes, so long builds and test runs count as work.
- Turn state from the transcript, read from the last entry's
stop_reason. A turn is over only onend_turn(orstop_sequence/max_tokens/refusal);tool_usemeans more is coming, so the session stays "working" even while the process sits quietly waiting on the API. This matters more than it sounds: assistant text and thinking blocks are logged mid-turn too, so judging by content alone marks a busy session finished dozens of times per turn, and runners keep dashing off and re-entering.
The 5-hour window is estimated the way ccusage does it — first activity after the previous block ended, floored to the hour, plus five hours. It can differ from the server's own reset by a few minutes.
Genuinely a two-minute PR. Sprites are drawn in code — pick one of six body
templates, give it a palette and a distinguishing feature, and you're done. It's
all in Sources/VibeCheck/Sprites.swift.
"🦔": Species(.small, body:(0.55,0.42,0.32), accent:(0.30,0.24,0.18), accessory:.spikes),Preview your work with ./build/VibeCheck --dump-sprites out.png. Requires macOS 13+
and the Xcode Command Line Tools.
한국어 README · MIT License