Talk to your phone.
It does the rest.
A voice-first AI agent for Android. Speak a task — Handy AI watches your screen, narrates each step aloud, and taps, types, and swipes through your apps until it's done. For everyone, and built with vision and motor accessibility in mind.
Designed for accessibility.
Other accessibility apps describe your screen. Handy AI does the task. Tell it what you want — send a message, order food, set a reminder — and it executes across any app on your phone. The Accessibility API was built for this; we're using it the way it was meant to be used.
Voice-first, end to end
Speak the task. Hear the agent's plan, every action it takes, and the result. Optional verbosity from brief to detailed. Default ON when TalkBack is detected.
TalkBack-compatible
The app's own UI works with TalkBack — every button has a proper label and the chat is announced as it arrives. Verified on Pixel.
Voice-confirm destructive actions
Before tapping Send, Buy, Delete, or other destructive buttons, Handy AI speaks a warning so you can hit Stop if the agent got it wrong.
Easy Mode
One giant mic button. No chat history, no menus, no settings to navigate. Designed for users with cognitive load, motor disabilities, or who simply want the simplest interface possible.
Set up by computer
Don't paste a 100-character API key on the phone. Get a 6-character code, enter it on your laptop at gethandyai.app/setup, paste your key with your desktop screen reader, and the phone receives it.
Bluetooth headsets + hearing aids
Audio routes through whatever output your phone uses — including Bluetooth headsets, hearing aids, and the spoken-feedback channel TalkBack already configured for you.
Voices from the beta.
This is where real beta testers' quotes go. We're not going to put words in anyone's mouth — this section stays empty until the first organic feedback comes in from blind, low-vision, motor-impaired, and abled testers actually using Handy AI.
No quotes yet.
Are you a beta tester? Reply to the welcome email with a one-sentence quote and the name + use case you'd want attributed (or "anonymous"). We'll post it here with your permission, exactly as you wrote it.
Describe it. Handy does it.
No macros, no workflows you have to build. Handy AI looks at your screen, one step at a time, decides the next tap, and keeps going until the task is done.
Voice or text
Tap the mic and say it. Or type. "Message Sarah on WhatsApp that I'm running late."
Sees what you see
Handy takes a screenshot, figures out the next tap, and executes. Then takes another. Until it's done.
Stops when you say
Big Stop button flips the agent off instantly. Won't enter passwords or run destructive actions unless you explicitly ask.
Join the Handy AI beta.
Drop your email and we'll send you install instructions plus the Play Store tester link as soon as you're on the list. Beta testers get a direct line to us and shape what ships next.
Install in 90 seconds.
Live on Play Store Closed Testing now; public Play Store listing rolling out as Google's review clears. For direct install, grab the signed APK - same bundle Google approved. Using a screen reader? Open gethandyai.app/setup on your computer to send the API key to your phone via a 6-character code.
Download
Allow install
Android will ask once — tap "Settings", toggle "Allow from this source", and back out.
Paste your key
Handy AI uses Claude. Paste your Anthropic key (get one at console.anthropic.com) — stored encrypted, on-device only.
Grant Accessibility
After your first send, Handy asks for Accessibility permission. That's how it taps on your behalf. Revoke any time.
Short answers.
Is it safe to install?
Yes — the APK is signed with our release key. It's the same bundle currently under Google Play review. Source is open at github.com/fainir/handy-ai.
What phones does it work on?
Android 11 and up. The Accessibility screenshot API requires API level 30+.
What does it cost?
The app is free. You pay Anthropic directly for the API calls Claude makes — usually a few cents per task. No subscription, no account.
Can it enter passwords or run payments?
The system prompt tells it not to, unless you explicitly ask. That's guidance to the model, not a hard sandbox — treat it accordingly.
Where does my screen content go?
To Anthropic's Claude API, one screenshot per step, only while a task is running. We never store, proxy, or train on it. See the privacy policy.