A voice for your Mac that never leaves it
Yubi (ゆび) lives in the menu bar. Hold the right Option key and talk, and the words land where the cursor is. Hold the right Command key and ask, and she reads the screen and clicks for you, after you say yes. Speech to text runs on the Mac. The app is 6 MB. I designed her and the site, wrote the Swift app, the Cloudflare Worker behind it and the billing, and shipped it as a direct download in September 2026
[ What it is ]- Role
- Founder, Product Designer and Engineer
- Timeline
- September to October 2026
- Scope
- Product design, Character and art direction, macOS engineering, Cloudflare Worker and metering, Billing
The challenge
The assistant I was using weighed 785 MB and sent my voice to the cloud
I wanted to talk to my Mac the way I talk to a colleague across the desk: say it once, have it typed in the app I am already in, and when I am lost, have someone point at the button. The tool I used for this shipped cloud speech recognition, screen recording, two analytics SDKs and a bundled command line agent, 785 MB in all. It guessed whether I wanted dictation or an action, and guessed wrong often enough that I stopped trusting it. And it spoke back in a synthesized voice that made every answer feel like a phone tree.
What I did
Two keys, one yes, and nothing synthesized
Right Option types what you say. Right Command asks her to help with the screen. Two keys instead of one guess, because a tool that has to infer your intent is a tool you have to watch. Speech recognition runs on the Mac with Apple's own engine, so no audio ever leaves it. The model call goes through a small Cloudflare Worker that holds the key, meters each task and streams the answer back; the app never carries a credential. Before anything risky, sending a message, pressing Return, opening a URL, she shows exactly what will happen and waits for your yes. And she never speaks generated text. Her voice is a set of recorded Japanese lines, so she sounds like one person, every time.
The one bit manga art is a design decision, not a style. A black and white sprite stays sharp at 48 pixels in the menu bar and at full size on the site, and it reads as drawn, not rendered. The same character now sits on a desk display and on a sheet of LINE stickers, which is the kind of thing that happens when you design the person before the product.
Outcomes
Hold a key, talk, let go. The words land where the cursor is, and nothing leaves the Mac
- Dictation into any app, unlimited and free, with speech recognition on the Mac. No audio is uploaded.
- Right Command reads the screen and clicks, types or opens things for you, and asks before anything it cannot undo.
- Sign in is a magic link. Free gives 10 tasks a month, Personal ¥1,480 for 60, Pro ¥2,980 for 150. Stripe inside the app.
- The app is 6 MB, notarized, and runs in English and Japanese.
- Character, site, Swift app, Worker, metering and billing, all designed, built and run by me.
Live at yubichan.ai, versions 0.6.0 to 0.8.1 shipped between September and October 2026.
