Live Platforms
Tappy
A voice assistant that lives in your MacBook's notch.
A native macOS assistant that turns speech into action on the Mac in front of you. Hold two keys, say what you want in ordinary words, and let go: apps open, windows arrange themselves, notes get written, selected text comes back rewritten. The panel grows out of the notch while you hold and shrinks back into it when it is done. Typical command: under a second.
Visit liveWhat is actually in it
Every figure is a count from the repository, not a projection.
Why it matters
Everyone has a version of the same friction. You know exactly what you want the machine to do, and doing it takes four steps with a trackpad. Find the window, drag it to half the screen, open the other two apps, copy the paragraph somewhere to be reworded. None of it is hard, which is the problem: it is too small to think about and too constant to ignore, and the interruption costs more than the task.
Tappy closes that gap by speaking. Hold two keys, say it in whatever words come out, and let go. The panel grows out of the notch while you hold and disappears when it is done, so there is no window to find and nothing to switch to. It matches meaning rather than phrasing, so there is no syntax to learn, and anything it did, it can undo.
The engineering is all in the middle. One model call decides four things at once, because asking them separately would mean four round trips on every command. Three quarters of commands skip the second call entirely, since "quit Spotify" needs no language understanding to resolve. The work itself runs through native macOS APIs rather than scripted clicks, so quitting an app takes 16 ms. That is how a spoken command finishes in a median of 1,285 ms, measured over 184 real ones rather than estimated.
Windows arranging themselves
See it work
Side by side
Tappy vs Siri, same job
Ask them both to open Chrome. Both clips start together and run at their own length, and the gap between them is the whole argument. Nothing sped up, nothing trimmed.
The problem
A Mac user's day is full of small mechanical tasks that break concentration without being difficult: finding the right window among fifteen and dragging it to half the screen, opening the four apps a piece of work needs one at a time, copying a paragraph somewhere to be reworded and pasting it back. None of these are hard. They are just constant, and the interruption outlasts the task. macOS already has Voice Control and Siri, but Voice Control is an accessibility tool built around describing the screen, and Siri has no spoken command for moving a window. The gap is an assistant that operates the machine rather than talking about it.
How it works
Hold two keys: Control and Option, from anywhere. No window to find and no click. The panel grows out of the notch while you hold, or drops out of the menu bar on a Mac without one.
Say it normally: not commands, just words. "Chuck some music on" and "play something" land in the same place, because it matches meaning rather than phrasing. There is no syntax to memorise.
Let go: it happens, usually before you have finished thinking about it. Every action reports what ran and how long it took, rather than leaving you to guess whether it worked.
Beyond the main loop
Apps and windows
Open an app, quit a named app, quit what you are in, close or hide everything, fullscreen, maximise, minimise, and snap a window to a half or a corner. "Open Spotify in fullscreen and put Chrome on the left half" is two instructions in one breath: it splits them, runs them in order, and waits for each app to actually be there before touching its window.
Notes and text
Make a note, type for you, or rewrite selected text in place. Rewriting works in any app: select text anywhere, say how you want it, and it comes back changed where it was. Say "format this" on a wall of text and it picks the structure itself, numbered for steps, bullets for parallel points, headings when there are several topics.
Web, music and system
Ask a question, search the web, or check what tab you are on. Control playback and ask what is playing. Set volume, flip dark or light mode, and lock the screen.
Undo that
Anything Tappy did, it can put back. Quitting every app is recoverable with two words, and anything with unsaved work asks first.
Routines
A sequence you define once and trigger by speaking. You describe the moment rather than the words, so "focus block", "going heads down" and "let me concentrate" all reach the same routine. Thirteen kinds of step, arranged by dragging, including running any macOS Shortcut you already have: anything you have already automated gets a spoken trigger without writing code.
Wake word, off by default
Say "hey tappy" and speak, hands free, from across the room. It ships off, because leaving a microphone open should be a decision rather than a surprise.
Stack
Decisions worth defending
One model call decides four things at once
Routing costs about 500 ms and that is the floor. A single call works out whether the speech was even addressed to the computer, whether it is one action or several, which capability handles it, and how hard it would be to undo. Asking those separately would mean four round trips for every command.
Extraction is skipped 74% of the time
"Quit Spotify" needs no second model call to work out what "Spotify" is. Only commands that genuinely need language understanding pay for one, which is why 136 of 184 measured commands finish in a median of 792 ms. Rewriting text is the honest exception at about 3.5 seconds, because a second model is writing prose. That one is worth the wait. The others would not be.
Native APIs, not scripted UI clicks
Actually doing the thing takes 2 ms to 300 ms. Quitting an app takes 16 ms and fullscreening a window takes 257 ms, because the work goes through macOS APIs rather than driving the interface the way a person would.
The app ships no API key
Requests go through a proxy that meters usage per user, so credit can be granted and revoked without shipping a new build, and nothing extractable from the binary is worth stealing.
Your voice never leaves the Mac
Speech is transcribed on-device. No audio is uploaded and none is saved to a file. What leaves is the text of the command and the name of the app you are in, which is what lets "put this on the left half" mean anything. No account, no tracking, no analytics, and exactly one file in the codebase makes network requests. On macOS 26 and later, rewriting can use Apple's on-device model instead, so selected text stops leaving the machine too.
It is drawn to match the notch it comes out of
The panel is the only part of Tappy you ever see, so it uses the timing and easing of native macOS animation, and ships eleven themes: solid, gradient and patterned fills with dot, grid, scanline and hatch motifs.
Outcome
- Median command 1,285 ms end to end, with the fastest quarter under 630 ms, measured over 184 real commands on a MacBook Air.
- Quitting an app 508 ms, minimising a window 674 ms, fullscreening 754 ms, opening an app 1,040 ms.
- Live in limited early access, released a fixed number of places at a time so every user can be answered.