Kumquat for macOS: a file converter triggered by a Shift gesture, no Accessibility permission required

⚡ The gist in 5 seconds

• Kumquat is an open-source (MIT) file converter that lives in the macOS menu bar: drag a file, tap Shift briefly, and a wheel of actions pops up under your cursor.
• Requirements: Apple Silicon and macOS 14 or newer, with an interface in Russian and English.
• The real trick isn't the conversion, it's the gesture detection: it works without Accessibility or Input Monitoring permissions.

🔍 What happened

An author going by TheAlexRay (kipperstudio) posted a breakdown of their app Kumquat on Habr — an open-source (MIT) file converter for macOS. The article appeared in the "From the sandbox" section: Habr. The set of actions is practical: "to JPG," "compress to 2 MB," "recognize text," "merge into PDF" — the result is saved next to the original, and the original itself is left untouched.

Source: Habr

Here's how the gesture works: start dragging files from Finder, the desktop, or your email, and while still holding the mouse button, briefly press Shift. A radial wheel unfolds under the cursor; hovering on a category opens it, hovering on the center goes back, and dropping onto a petal triggers the action. The wheel only shows options that fit all the dragged files at once: drag a HEIC and a MOV together, and there's no "to MP3" option.

The most interesting part is how the gesture is caught. Instead of a global CGEventTap or NSEvent.addGlobalMonitorForEvents, which require Accessibility or Input Monitoring, the author polls public system state: NSEvent.pressedMouseButtons, NSEvent.modifierFlags, and NSPasteboard(name:.drag). A timer ticks once every 120 ms while the button is released and switches to 60 Hz as soon as the button is pressed. The core is DragGestureMonitor.swift, and the key line from it: let mouseDown = NSEvent.pressedMouseButtons & 1!= 0. Buried inside, it turns out that ffmpeg and nine codecs had to be built from source, and most of the code was written by Claude under the author's supervision.

🧠 Our take

Behind the news is an everyday pain point: an iPhone photo needs to be converted to a format a website understands, a screen recording needs to be squeezed under an email attachment limit, a screenshot of an error needs its text extracted. Every time, that means either a paid "all-in-one" app or a web service where you upload your personal files. A local menu bar utility is a logical answer, and the fact that one person built it with an AI assistant is more of a backdrop than the point here.

Our assessment: Kumquat's main contribution to the ecosystem isn't the converter, it's the permission-free drag detection. Apps that require Accessibility get access to every keystroke in other windows, and the user physically cannot verify what else they're listening to. Polling pressedMouseButtons and the drag pasteboard gives a narrow slice of state. The cost is a constantly running timer — and that's where the battery debate begins.

We believe Mac users with personal files win, while web converters and subscription utility bundles lose. But the Apple Silicon-only requirement cuts off Intel Macs, and there's not a word about code signing or notarization in the post — for a utility that touches personal documents, that's no small detail.

⚠️ What we don't know

There are no numbers: the author promises to explain how much faster VideoToolbox is than libx264, but no concrete benchmarks appear in the post. It doesn't say what OCR was built with, how the app is signed and notarized, how much energy the 60 Hz polling consumes, or how the wheel behaves in sandboxed apps. There's a single source — the author's article about their own project — with no independent testing. There's a separate section about Claude's mistakes that we haven't read, so don't take it on faith: it's worth verifying yourself on your own files.

Prediction: within a quarter, at least one fork or clone with the same permission-free drag detection will appear — the technique is described publicly, the license is MIT, and the code is only about a hundred lines. If no forks show up, it means the approach doesn't transfer well beyond this particular task.

Disclaimer: a single source — the author's article about their own project; nothing has been officially confirmed.

Source: Habr