Open-AutoGLM

中文阅读.

👋 Join our WeChat community

Project Introduction

Phone Agent is a mobile intelligent assistant framework built on AutoGLM. It can understand phone screen content in a multimodal way and help users complete tasks through automated operations. The system controls devices through ADB ( Android Debug Bridge), uses vision-language models for screen perception, and combines intelligent planning capabilities to generate and execute operation workflows. Users only need to describe their requirements in natural language, such as "Open Xiaohongshu and search for food", and Phone Agent will automatically parse the intent, understand the current interface, plan the next action, and complete the entire workflow. The system also has a built-in sensitive operation confirmation mechanism and supports manual takeover in login or verification code scenarios. Additionally, it provides remote ADB debugging capabilities, allowing device connection via WiFi or network for flexible remote control and development.

⚠️ This project is for research and learning purposes only. It is strictly prohibited to use it for illegal information gathering, system interference, or any illegal activities. Please carefully review the Terms of Use.

Model Download Links

Model	Download Links
AutoGLM-Phone-9B	🤗 Hugging Face 🤖 ModelScope

Currently, this model only supports Chinese. A multilingual version of the model is coming soon.

Environment Setup

1. Python Environment

Python 3.10 or above is recommended.

2. ADB (Android Debug Bridge)

1. Download the official ADB installation package, and extract it to a custom path
2. Configure environment variables

• MacOS configuration method: In Terminal or any command-line tool

  # Assuming the extracted directory is ~/Downloads/platform-tools. If not, please adjust the command accordingly.
  export PATH=${PATH}:~/Downloads/platform-tools
  

• Windows configuration method: You can refer to this third-party tutorial for configuration.

3. Android 7.0+ Device or Emulator with Developer Mode and USB Debugging Enabled

1. Enable Developer Mode: The typical method is to find Settings > About Phone > Build Number and tap it quickly about 10 times until a popup shows "Developer mode enabled". Different phones may vary slightly; if you can't find it, search online for a tutorial.
2. Enable USB Debugging: After enabling Developer Mode, go to Settings > Developer Options > USB Debugging and enable it
3. Some devices may require a restart after enabling developer options for changes to take effect. You can test it: connect your phone to the computer via USB cable, then run adb devices to check if device information appears. If not, the connection has failed.

4. Install ADB Keyboard (for text input)

Download the installation package and install it on the corresponding Android device. Note: After installation, you need to enable ADB Keyboard in Settings > Input Method or Settings > Keyboard List for it to work.

Deployment Preparation

1. Install Dependencies

pip install -r requirements.txt 
pip install -e .

2. Configure ADB

Make sure your USB cable supports data transfer, not just charging.

Ensure ADB is installed and connect the device with a USB cable:

# Check connected devices
adb devices

# Output should show your device, e.g.:
# List of devices attached
# emulator-5554   device

3. Start the Model Service

1. Download the model and install the inference engine framework according to the For Model Deployment section in requirements.txt.
2. Start via SGlang / vLLM to get an OpenAI-compatible service. Here's a vLLM deployment solution, please strictly follow our provided startup parameters:

• vLLM:

python3 -m vllm.entrypoints.openai.api_server \
 --served-model-name autoglm-phone-9b \
 --allowed-local-media-path /   \
 --mm-encoder-tp-mode data \
 --mm_processor_cache_type shm \
 --mm_processor_kwargs "{\"max_pixels\":5000000}" \
 --max-model-len 25480  \
 --chat-template-content-format string \
 --limit-mm-per-prompt "{\"image\":10}" \
 --model zai-org/AutoGLM-Phone-9B \
 --port 8000

• This model has the same architecture as GLM-4.1V-9B-Thinking. For detailed information about model deployment, you can also check GLM-V for model deployment and usage guides.

• After successful startup, you can access the model service via http://localhost:8000/v1. If you deploy the model on a remote server, use that server's IP to access the model.

Using AutoGLM

Command Line

Set the --base-url and --model parameters according to your deployed model. For example:

# Interactive mode
python main.py --base-url http://localhost:8000/v1 --model "autoglm-phone-9b"

# Specify model endpoint
python main.py --base-url http://localhost:8000/v1 "Find the top-rated cinema nearby and navigate me to there by foot"

# List supported apps
python main.py --list-apps

Python API

from phone_agent import PhoneAgent
from phone_agent.model import ModelConfig

# Configure model
model_config = ModelConfig(
    base_url="http://localhost:8000/v1",
    model_name="autoglm-phone-9b",
)

# Create Agent
agent = PhoneAgent(model_config=model_config)

# Execute task
result = agent.run("Open Taobao and search for wireless earphones")
print(result)

Remote Debugging

Phone Agent supports remote ADB debugging via WiFi/network, allowing device control without USB connection.

Configure Remote Debugging

Enable Wireless Debugging on Phone

Make sure the phone and computer are on the same WiFi network, as shown below:

Use Standard ADB Commands on Computer

# Connect via WiFi, change to the IP address and port shown on your phone
adb connect 192.168.1.100:5555

# Verify connection
adb devices
# Should show: 192.168.1.100:5555    device

Device Management Commands

# List all connected devices
adb devices

# Connect to remote device
adb connect 192.168.1.100:5555

# Disconnect specific device
adb disconnect 192.168.1.100:5555

# Execute task on specific device
python main.py --device-id 192.168.1.100:5555 --base-url http://localhost:8000/v1 --model "autoglm-phone-9b" "Open Douyin and browse videos"

Python API Remote Connection

from phone_agent.adb import ADBConnection, list_devices

# Create connection manager
conn = ADBConnection()

# Connect to remote device
success, message = conn.connect("192.168.1.100:5555")
print(f"Connection status: {message}")

# List connected devices
devices = list_devices()
for device in devices:
    print(f"{device.device_id} - {device.connection_type.value}")

# Enable TCP/IP on USB device
success, message = conn.enable_tcpip(5555)
ip = conn.get_device_ip()
print(f"Device IP: {ip}")

# Disconnect
conn.disconnect("192.168.1.100:5555")

Remote Connection Troubleshooting

Connection refused:

• Ensure device and computer are on the same network
• Check if firewall is blocking port 5555
• Confirm TCP/IP mode is enabled: adb tcpip 5555

Connection dropped:

• WiFi may have disconnected, use --connect to reconnect
• Some devices disable TCP/IP after restart, requiring USB to re-enable

Multiple devices:

• Use --device-id to specify which device to use
• Or use --list-devices to view all connected devices

Configuration

Custom SYSTEM PROMPT

Directly modify the configuration file phone_agent/config/prompts.py

1. You can inject system prompts to enhance the model's capabilities in specific domains
2. You can inject app names to disable certain apps

Environment Variables

Variable	Description	Default
PHONE_AGENT_BASE_URL	Model API address	http://localhost:8000/v1
PHONE_AGENT_MODEL	Model name	autoglm-phone-9b
PHONE_AGENT_MAX_STEPS	Max steps per task	100
PHONE_AGENT_DEVICE_ID	ADB device ID	(auto-detect)

Model Configuration

from phone_agent.model import ModelConfig

config = ModelConfig(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",  # API key (if required)
    model_name="autoglm-phone-9b",  # Model name
    max_tokens=3000,  # Maximum output tokens
    temperature=0.1,  # Sampling temperature
    frequency_penalty=0.2,  # Frequency penalty
)

Agent Configuration

from phone_agent.agent import AgentConfig

config = AgentConfig(
    max_steps=100,  # Max steps per task
    device_id=None,  # ADB device ID (None for auto-detect)
    verbose=True,  # Print debug info (including thinking process and actions)
)

Verbose Mode Output

When verbose=True, the Agent outputs detailed information at each step:

==================================================
💭 Thinking Process:
--------------------------------------------------
Currently on the system home screen, need to launch the Xiaohongshu app first
--------------------------------------------------
🎯 Executing Action:
{
  "_metadata": "do",
  "action": "Launch",
  "app": "Xiaohongshu"
}
==================================================

... (continue to next step after executing action)

==================================================
💭 Thinking Process:
--------------------------------------------------
Xiaohongshu is open, now need to tap the search box
--------------------------------------------------
🎯 Executing Action:
{
  "_metadata": "do",
  "action": "Tap",
  "element": [500, 100]
}
==================================================

🎉 ================================================
✅ Task Complete: Successfully searched for food guides
==================================================

This allows you to clearly see the AI's reasoning process and specific operations at each step.

Supported Apps

Phone Agent supports 50+ mainstream Chinese apps:

Category	Apps
Social & Chat	WeChat, QQ, Weibo
E-commerce	Taobao, JD.com, Pinduoduo
Food & Delivery	Meituan, Ele.me, KFC
Travel	Ctrip, 12306, Didi
Video & Entertainment	Bilibili, Douyin, iQiyi
Music & Audio	NetEase Music, QQ Music, Ximalaya
Life Services	Dianping, Amap, Baidu Maps
Content Communities	Xiaohongshu, Zhihu, Douban

Run python main.py --list-apps to see the complete list.

Available Actions

The Agent can perform the following actions:

Action	Description
Launch	Launch an app
Tap	Tap at specified coordinates
Type	Input text
Swipe	Swipe the screen
Back	Go back to previous page
Home	Return to home screen
Long Press	Long press
Double Tap	Double tap
Wait	Wait for page to load
Take_over	Request manual takeover (login/captcha)

Custom Callbacks

Handle sensitive operation confirmation and manual takeover:

def my_confirmation(message: str) -> bool:
    """Sensitive operation confirmation callback"""
    return input(f"Confirm execution of {message}? (y/n): ").lower() == "y"


def my_takeover(message: str) -> None:
    """Manual takeover callback"""
    print(f"Please complete manually: {message}")
    input("Press Enter to continue after completion...")


agent = PhoneAgent(
    confirmation_callback=my_confirmation,
    takeover_callback=my_takeover,
)

Examples

Check the examples/ directory for more usage examples:

• basic_usage.py - Basic task execution
• Single-step debugging mode
• Batch task execution
• Custom callbacks

Development

Configure Development Environment

Development requires dev dependencies:

pip install -e ".[dev]"

Run Tests

pytest tests/

Complete Project Structure

phone_agent/
├── __init__.py          # Package exports
├── agent.py             # PhoneAgent main class
├── adb/                 # ADB utilities
│   ├── connection.py    # Remote/local connection management
│   ├── screenshot.py    # Screen capture
│   ├── input.py         # Text input (ADB Keyboard)
│   └── device.py        # Device control (tap, swipe, etc.)
├── actions/             # Action handling
│   └── handler.py       # Action executor
├── config/              # Configuration
│   ├── apps.py          # Supported app mappings
│   └── prompts.py       # System prompts
└── model/               # AI model client
    └── client.py        # OpenAI-compatible client

FAQ

We've listed some common issues and their solutions:

Device Not Found

Try resolving by restarting the ADB service:

adb kill-server
adb start-server
adb devices

Text Input Not Working

1. Ensure ADB Keyboard is installed on the device
2. Enable it in Settings > System > Language & Input > Virtual Keyboard
3. The Agent will automatically switch to ADB Keyboard when input is needed

Screenshot Failed (Black Screen)

This usually means the app is displaying a sensitive page (payment, password, banking apps). The Agent will automatically detect this and request manual takeover.

Citation

If you find our work helpful, please cite the following paper:

@article{liu2024autoglm,
  title={Autoglm: Autonomous foundation agents for guis},
  author={Liu, Xiao and Qin, Bo and Liang, Dongzhu and Dong, Guang and Lai, Hanyu and Zhang, Hanchen and Zhao, Hanlin and Iong, Iat Long and Sun, Jiadai and Wang, Jiaqi and others},
  journal={arXiv preprint arXiv:2411.00820},
  year={2024}
}
@article{xu2025mobilerl,
  title={MobileRL: Online Agentic Reinforcement Learning for Mobile GUI Agents},
  author={Xu, Yifan and Liu, Xiao and Liu, Xinghan and Fu, Jiaqi and Zhang, Hanchen and Jing, Bohao and Zhang, Shudan and Wang, Yuting and Zhao, Wenyi and Dong, Yuxiao},
  journal={arXiv preprint arXiv:2509.18119},
  year={2025}
}