Just say it.
Your phone
does it.
OpenPhoneUse is an AI agent that sees your Android screen, taps, types and moves between apps, right on the phone. No PC. No root.
- Open Clock
- Expand the 6:45 alarm
- Repeat Mon – Fri
- Check it stuck
It reads the screen like you do.
Every tappable thing gets a number. Android's accessibility tree gives the exact labels; a marked screenshot covers icons, maps and games.
It makes a plan first.
Your words become a short plan. The panel at the bottom shows each move while it works, and Stop is always one tap away.
Then it just does it.
It taps, types, scrolls and switches apps through the accessibility service, and waits for the screen to settle before the next move.
And checks before it says done.
It reads the end state back off the screen, then tells you exactly what it did.
Not a macro. Not a script. An agent that reads any app the way you do, and stops to ask before anything that costs you.
It asks first. Every time it matters.
These are the app's real cards, in light and dark. Text on web pages and in other apps is read as data, never as orders. Go on, press the buttons.
No PC. No root. No ADB.
Other phone agents need a computer plugged in, or a hacked phone. This one needs a switch you flip yourself.
Any model you already pay for.
Sign in with your ChatGPT plan, or paste a key for OpenAI, Claude, Gemini or Grok. Switch whenever you like, even mid-conversation. GPT-6.1 Terra with fast mode is the default.
It looks, so it's 2.3× faster.
A marked screenshot alongside the text tree. Same model, same tasks, against text only.
No server. No analytics.
Your screen goes straight to the model you chose, and only while a task needs it. Keys stay in Android Keystore. Nothing of ours sits in between, because there is nothing of ours.
31 real tasks. Graded blind.
Every task starts from the same phone snapshot. Rubrics check the device, the answer and the steps; a judge model scores the rest from the screenshots, without knowing which model ran.
From the 2026-11-03 run with GPT-6.1 Terra in fast mode.
Ask. Then watch.
Built in the open.
Every line of code that touches your screen is on GitHub. Read it, fork it, make it better.
Questions, answered.
What is OpenPhoneUse?
An open-source Android app that lets an AI agent operate your phone like a person. It reads the screen through Android's AccessibilityService, looks at a screenshot, and taps, types, scrolls and switches apps to finish what you ask in plain words.
Does it need root, ADB, Shizuku or a PC?
No. It uses only Android's public accessibility API, which you turn on yourself in Settings › Accessibility.
Which AI models does it work with?
Your ChatGPT plan through OpenAI's official Sign in with ChatGPT, an OpenAI API key, Claude, Grok and Gemini. GPT-6.1 Terra with fast mode is the default.
Is it safe to let it use my phone?
Payments and security changes always wait for your tap. Deletes and sends go through only when you named exactly what. The model never types passwords, and text from web pages and other apps is treated as data, never as instructions.
Does my screen go to the cloud?
Only to the model provider you picked, and only while a task needs the screen. There is no OpenPhoneUse server and no analytics.
Can I stop it mid-task?
Yes. The panel at the bottom of the screen has a Stop button the whole time it works, and the screen is yours again the moment you press it.
Give your phone a pair of hands.
Free. MIT-licensed. Bring your own model.