Mobile Use 📱

Mobile Use​: Your AI assistant for mobile - Any app, any task.

[ English | 中文 ]

TurnOnBluetoothAndWIFI_en_2x.mp4

The user inputs the task description on the Web interface, and the Mobile Use automatically operates the mobile phone and completes the task.

📊 Benchmark

In the AndroidWord dynamic evaluation environment, we evaluated the Mobile Use agent solution with the multimodal large language model Qwen2.5-VL-72B-Instruct and achieved a 38% success rate.

✨ Key Features

• Auto-operating the phone: Automatically operate the UI to complete tasks based on user input descriptions.
• Smart Element Recognition: Automatically parses GUI layouts and identifies operational targets.
• Complex Task Processing: Supports decomposition of complex task and multi-step operations.

🚀 Quick Start

mobile-use requires ADB to control the phone, which necessitates the prior installation of the relevant tools and connecting the phone to the computer via USB.

1. Install SDK Platform-Tools

• Step 1. Download SDK Platform-Tools for Desktop, click there.

• Step 2. Unzip the downloaded file and add the platform-tools path to the environment variables.

  • Windows

    In Windows, you can add the platform-tools PATH to the Path environment variable on the graphical interface (see here) or through the command line as follows:

    setx PATH "%PATH%;D:\your\download\path\platform-tools"
    

  • Mac/Linux

    $ echo 'export PATH=/your/downloads/path/platform-tools:$PATH' >> ~/.bashrc
    $ source ~/.bashrc
    

• Step 3. Open the command line and enter adb devices (Windows: adb.exe devices) to verify adb is available or not.

2. Enable developer mode and open USB debugging on your phone

3. Connect your computer and phone using a USB cable. And verify the adb is connected.

Run the adb devices (Windows: adb.exe devices) command on the command line terminal. If the device serial_no is listed, the connection is successful. The correct log is as follows:

List of devices attached
a22d0110        device

4: Install mobile-use

With pip (Python>=3.10):

pip install mobile-use

5. Launch the webui service

python -m mobile_use.webui

6. Usage

Once the service starts successfully, open the address http://127.0.0.1:7860 in your browser to access the WebUI page, as shown below:

Click VLM Configuration to set the Base URL and API Key of the multimodal large language model. It is recommended to use the multimodal large language model of Qwen2.5-VL series.

Input task descriptions in the input box at the lower left corner, click start to execute tasks.

⚠️ Special Reminder: The actions are autonomously decided by the intelligent agent, which may pose uncontrollable operational risks. It is recommended that during the experience, you constantly monitor your phone's status. If you encounter any operational risks, promptly terminate the task or use a test phone for the experience to avoid issues caused by accidental operations.

🎉 More Demo

Case1：Search the latest news of DeepSeek-R2 in Xiaohongshu APP and forward one of the news to the Weibo App

search_forward_2x.mp4

Case2：Order 2 Luckin coffees with Meituan, 1 hot raw coconut latte standard sweet, and 1 cold light jasmine

order_coffee_en_2x.mp4

Case3：用美团点一杯咖啡，冰的，标准糖

demo01_2x.mp4

Case4：用美团帮我点2杯瑞幸咖啡，要生椰拿铁标准糖、热的

order_coffee_zh_2x.mp4

Case5：在浏览器找一张OPPO Find N5图片，询问DeepSeek应用该手机介绍信息，将找到的图片和介绍信息通过小红书发布

demo03_2x.mp4

Case6：帮我去OPPO商城、京东、以及淘宝分别看一下oppofind n5售价是多少

oppofindn5_price_zh_2x.mp4

⚙️ Advance

Advance Settings

📱 Mobile Settings

The Android ADB Server Host and Android ADB Server Port allow you to specify the address and port of the android ADB service, which can be used for remote device connections or local android ADB services on non-default port. When multiple devices exist, you need to specify the Device Serial No. The Reset to HOME parameter indicates whether to return the phone to the home page before executing the task. If you continue the previous task, you need to cancel this option.

⚙️ Agent Settings

The Max Run Steps parameter specifies the maximum number of iteration steps for the Agent. If the current task exceeds the maximum number of iteration steps, the task will be stopped. Therefore, you are advised to set a larger value for complex tasks with more operation steps. The Maximum Latest Screenshot is to control the number of latest screenshots that the Agent can see. Because pictures consume more tokens, when the task has more steps, Appropriately take a Screenshot of the latest Maximum Latest Screenshot and send it to VLM to generate the next operation accordingly. The Maximum Reflection Action is to control the maximum number of reflection times of the Agent. The greater the value, the higher the fault tolerance rate of the Agent, but the longer the processing time of the task.

🔧 VLM Configuration

Click VLM Configuration to specify the Base URL and API Key of the multimodal large language model, as well as the model name and temperature coefficient. It is recommended to use the multimodal large language model of Qwen2.5-VL series.

Use agent with code

import os
from dotenv import load_dotenv
from mobile_use.scheme import AgentState
from mobile_use import Environment, VLMWrapper, Agent
from mobile_use.logger import setup_logger

load_dotenv()
setup_logger(name='mobile_use')

# Create environment controller
env = Environment(serial_no='a22d0110')
vlm = VLMWrapper(
    model_name="qwen2.5-vl-72b-instruct", 
    api_key=os.getenv('VLM_API_KEY'),
    base_url=os.getenv('VLM_BASE_URL'),
    max_tokens=128,
    max_retry=1,
    temperature=0.0
)

agent = Agent.from_params(dict(type='default', env=env, vlm=vlm, max_steps=3))

going = True
input_content = goal
while going:
    going = False
    for step_data in agent.iter_run(input_content=input_content):
        print(step_data.action, step_data.thought)

🗺️ Roadmap

• Improve agent memory and reflection (summarize, compress.)
• Provide multi-agent implementation
• Provide an evaluation process about AndroidWorld dynamic environment
• Develop an APP that can be installed directly on the phone

🌱Contributing

We welcome all forms of contributions! Please read our contribution guide to learn about:

• How to submit an issue to report problems.
• The process of participating in feature development, See Developer Document.
• Code style and quality standards, See Developer Document.
• Methods for suggesting documentation improvements.

📜 License

This project is licensed under the MIT License, which permits free use and modification of the code but requires retaining the original copyright notice.

📚 Citation

If you have used this project in your research or work, please cite:

@software{
  title = {Mobile Use​: Your AI assistant for mobile - Any app, any task},
  author = {Jiamu Zhou, Xiaoyun Mo, Ning Li, Qiuying Peng},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/MadeAgents/mobile-use}
}

🤝 Acknowledgements

This project benefits from the contributions of:

• Inspiration from browser-use
• The multimodal large language model for the agent is based on Qwen2.5-VL
• The Web UI is built on Gradio

Thanks for their wonderful works.