aibar is an AI assistant living in your MacBook's notch. It hears you through a speech model running on the Neural Engine and answers through the model you choose: Apple's on-device intelligence, your own local server, or a Gemini key. It also operates the Mac itself: volume, brightness, Wi-Fi, timers, apps.
250–400 ms
speech becomes text on an M1, entirely on-device
15
system tools the assistant can actually operate
<50 MB
resident memory while it waits in the notch
24 h
on battery with aibar running, measured against a control day without it. Its draw came out smaller than the difference between two ordinary days, too small to separate from the machine's own noise.
aibar's tools report FAILED when an action doesn't happen, and the model isn't allowed to call it done. This page follows the same rule, which is why the battery figure sat empty here for weeks until someone actually measured it.
The measurement: one MacBook Air M1 on battery for a full working day with aibar running in the background, against the same machine on an ordinary day without it. A working Mac spends its battery on the work: the browser, the editor, the compiler. The question was only whether aibar adds to that. It didn't. The day with aibar actually ran slightly cheaper than the day without, which is not a claim that it saves power; it means the difference is smaller than the gap between any two ordinary days.
One thing did show up. With the lid shut, aibar cost about 0.4 % an hour against a 0.1 % floor, roughly three percent of a charge overnight. Small in absolute terms, and it would have been easy to leave out. The cause was two background pollers that woke the machine every three seconds and had no reason to run while the screen was asleep; they now stop when it does. The fix was then measured the same way it was found: a second overnight run on the patched build put the sleeping cost back at the 0.1 % floor. That is indistinguishable from the same Mac without aibar.
Speech recognition is Parakeet v3 on the Neural Engine: English and Russian in one model, roughly a third of a second to text. The reply appears as text in the notch. No audio is streamed anywhere.
Hold Fn, speak, release. The words type themselves into whatever app has focus.
⌘⇧Space starts a voice query without leaving what you're doing. A red pill in the notch shows the mic is live.
Drag files into the notch, carry them between apps and Spaces, drop them where they're needed.
Weather, now playing, CPU, calendar, clipboard: a glanceable strip that only polls while open.
⌘⇧1, drag a rectangle over anything: a PDF, an error dialog, a chart. The Vision framework reads the text on this Mac, the screenshot is deleted, and the text goes into your next question. The image never leaves the machine; the text goes wherever your model runs: nowhere with Apple on-device or a local server, to Google with a Gemini key.
Volume, brightness, dark mode, Wi-Fi, Bluetooth, music, timers, opening apps, running Shortcuts. When a tool fails, it reports FAILED and the assistant says so. Never a false "done".
Apple's on-device model: free, private, English. Any OpenAI-compatible local server: Ollama, LM Studio, llama.cpp. Or Gemini with a free API key, for cloud answers in any language.
NotchNook proved the notch is worth using, and it's a polished product. The category's real cost is what runs while nobody is looking: in July 2026 the maintainer of Boring Notch traced a battery-drain report to a bug in its music visualizer, and this September its visualizer was reworked again to cut compositor load. Pretty things in the notch are paid for in watts.
aibar's first engineering rule is that the notch must cost almost nothing while idle: models unload on battery, polling stops when the notch closes, pollers stop again when the screen sleeps, and there is no visualizer to animate. The day-long measurement above is what that rule buys: a full day where the battery could not tell whether aibar was running.
Siri AI shipped with macOS 27 on September 14, 2026, and it runs on any Apple-silicon Mac. The M3 and 12 GB floor applies only to its new voices and dictation. It is good, and it is Apple's. It also answers through Apple's cloud, is an English-only beta for now, and needs macOS 27. aibar stays on macOS 15+, keeps speech on the Neural Engine, understands Russian, and lets you pick a model that never leaves the room.
Yes, all of it during the beta. The long-term plan is a free hub with a paid Pro tier.
Speech recognition always stays on the Neural Engine. The language model is your call: Apple on-device and local servers never touch the network; Gemini is a cloud model behind your own key. System tools run on this Mac; the one exception is the weather tile, which asks MET Norway for the forecast in the city you typed. aibar never reads your location.
aibar has no backend. There is nothing to sign into, no account, and your API keys live in the macOS Keychain on your machine.
An assistant that changes your volume, toggles Wi-Fi and runs Shortcuts can't live inside the App Store sandbox. aibar ships instead as a notarized DMG signed with a Developer ID, and Gatekeeper checks it like any other Mac app.
macOS 15 or newer on Apple silicon. The Apple on-device model additionally needs macOS 26 with Apple Intelligence enabled; local servers and Gemini work everywhere.