austinintelligence/android-use

安卓应用 给AI一个安卓设备。 你的客服只需每个任务调用两次工具，就能看到屏幕、点击、输入、打开应用，并运行 Chrome。

安装· 联系代理人· 模拟器· 文献资料· 更新日志

一通电话，然后才改变了什么

在Android模拟器上，通过MCP进行真实会话。代理发送一个命令;结果就是它指向的界面。

画面是当屏幕上文字较少时，代理得到的画面：17 KB，每个控件都有编号框，所以tap e4是有效的。

减少往返次数和更短的结果才是关键。在基于 MCP 的 API 36 模拟器上测量，13 个任务各运行 10 次，对比 1.1 版本变更前的相同任务（方法和数字）：

分流延迟保持不变（p50大约700毫秒）。

android-use Give your AI agent an Android phone. Keep your hands on the leash. Your agent opens apps, reads the screen, taps, types, and drives Chrome on a real phone or an emulator, in about two tool calls per task. You decide what it may touch.

Install · Connect an agent · Control · Emulators · Documentation · Changelog

What is android-use?

Your agent lives in a terminal and a browser. Most of the apps people use every day live on a phone. android-use connects the two: it gives the agent you already use (Claude Code, Codex, Cursor, Gemini CLI, and others) hands on a real Android phone or emulator.

There are two pieces. A small helper app on the device reads the screen through Android’s accessibility service and carries out actions. au, on your computer, talks to the helper over ADB and gives your agent two tools: one to look, one to act. The agent asks in plain commands like open app "Settings" then tap "Network & internet" and gets back a short answer: the screen it reached, and after later actions, only what changed.

• It reads text first. The screen comes back as a list of what is there: labels, buttons, switches with their state (sw Wi-Fi=on), and fields with their contents. Agents act on labels, not on guessed coordinates. When an app shows little text, a small screenshot with a numbered box per control comes along, so tap e4 still works.
• It checks itself. Every action is matched against the screen it was planned from, journaled so it runs at most once, and confirmed after. A locked phone, a screen that is off, or an ambiguous label stops it with the reason and the one step that fixes it, instead of a blind tap.
• It remembers what worked. Save a command that worked as a routine and run it again in one call. Every session leaves a trace you can review, export, or turn into a test.