Give AI an Android device.
Plug in a phone or tablet. Your AI can see the screen, tap, type, open apps, and use Chrome on the real device.
Get started · Documentation · Examples · Agent setup · CLI reference
Android Use connects the AI agent on your computer to the Android device in your hand. Ask normally:
- “Open Settings and check the battery.”
- “Search for Yellowstone in Chrome and show me what you find.”
- “Open this app and go to the account screen.”
- “Tell me which Wi-Fi network is connected.”
The agent reads the same labels and controls you see, then taps, types, scrolls, and checks its work.
Real device, real result. Android Use opened this National Park Service page in Chrome on a connected Android tablet and confirmed that Yellowstone loaded.
You: Open Chrome, find Yellowstone, and show me the page.
Agent: Opened Chrome, loaded Yellowstone, and confirmed the page.
| Capability | What that means | |
|---|---|---|
| 👁 | See what is on screen | Understand visible text, buttons, fields, lists, and menus. |
| 👆 | Use your apps | Tap, type, scroll, go back, return home, and open apps. |
| 🌐 | Browse with Chrome | Open pages, follow links, fill fields, read results, and take page screenshots. |
| 📱 | Check the device | Read supported device information, location, and notifications when you allow it. |
| 📷 | Capture when you ask | Take screenshots or use supported camera, microphone, and screen recording with Android's permission. |
| 🔌 | Work with your agent | Connect Codex, Claude Code, Cursor, or another MCP-compatible agent. |
See the verified capability details →
You need Android 8 or newer, a data-capable USB cable, and a supported release host: Windows x86_64, macOS Apple Silicon, or Linux x86_64.
- Download the archive for your computer from Releases and unzip it.
- Connect and unlock your Android device.
- Open a terminal in the unzipped folder, then run:
# Windows PowerShell
.\au.exe setup# macOS or Linux
./au setupAndroid Use checks the cable, remembers the device, installs its small helper, and opens the right Android permission screen. Approve the prompts on the device and leave the command running. It finishes when everything is ready.
Android Use is ready
The release archive contains au and the Android Use helper together. Keep them in the same folder; the setup command installs the matching helper automatically.
Full human quickstart → · Troubleshooting →
If you want to use Android Use in Codex, Cursor, Claude Code, OpenClaw, Hermes, or another coding agent, paste this prompt into the agent:
Set up Android Use from the official release or repository. Preserve unrelated files, credentials, enrollments, and agent settings. Keep au beside aubridge.apk. Run au setup with one unlocked, USB-debugging-authorized device; pause for Android's Accessibility approval and resume with au doctor. Configure a local stdio MCP server using the absolute au path and serve --mcp. Verify with android.read command status and android.read command screen. Use the two command-string tools for normal work. Never use raw ADB, bypass Android prompts, or replay a partial or unknown mutation.
Then ask: “Open Settings and tell me which Wi-Fi network is connected.”
The full fallback runbook and manual commands are in the Agent installation and recovery guide.
If your client supports MCP, configure it to run:
au serve --mcpThe agent receives two focused tools:
android.readsees what is happening.android.actperforms the next small action.
Coding agents can start with AGENTS.md. It contains the whole operating loop and recovery rules in one page. For Codex, Claude Code, Cursor, and similar clients, see the Agent Quickstart.
Normal control stays between your computer and the connected device. Android still owns every sensitive permission prompt, and the helper cannot use the internet on its own. Read the security model →
| I want to… | Start here |
|---|---|
| Let Codex, Claude Code, or Cursor use my Android device | Agent Quickstart |
| Try commands myself | Human Quickstart |
| Build an agent integration | Agent protocol |
| Automate a repeatable task | Examples and JSONL |
| Understand permissions and remote-control risk | Security |
| Contribute to Android Use | Development |
Does Android Use include an AI model?
No. It gives an agent a safe, compact Android interface. Bring the agent or application you already use.
Does it work without screenshots?
Usually. Semantic UI is the preferred first read because it is smaller and identifies actionable controls. Use a screenshot when visual layout or imagery matters.
Does the phone need to be rooted?
No. Android Use relies on Android's normal debugging and user-granted accessibility or capability permissions.
Can it control any connected phone?
No. A server session is bound to one enrolled hardware identity. Android also asks you to trust the computer and separately controls sensitive permissions.
Can I use Wi-Fi instead of USB?
The current product enrolls an ADB endpoint, but release onboarding and validation use USB. Treat wireless ADB as an advanced Android transport and secure it as carefully as physical debugging access.
Where do captures and logs go?
Large captures are stored as local private artifacts. The operation journal stores bounded operation metadata, not screenshot or media contents. See Security for paths, retention, and removal.
Android Use 1.0.0 is the supported release line. The typed interface, bounded execution model, helper authentication, Chrome control, and local artifact system are implemented and tested. Optional Android capabilities vary by device; au doctor is the authority for the connected setup.
MIT licensed. Contributions are welcome—start with CONTRIBUTING.md.