Open Autoglm Phone Agent

Open-AutoGLM Phone Agent

Skill by ara.so — Daily 2026 Skills collection.

Open-AutoGLM is an open-source AI phone agent framework that enables natural language control of Android, HarmonyOS NEXT, and iOS devices. It uses the AutoGLM vision-language model (9B parameters) to perceive screen content and execute multi-step tasks like "open Meituan and search for nearby hot pot restaurants."

Architecture Overview

User Natural Language → AutoGLM VLM → Screen Perception → ADB/HDC/WebDriverAgent → Device Actions

• Model: AutoGLM-Phone-9B (Chinese-optimized) or AutoGLM-Phone-9B-Multilingual
• Device control: ADB (Android), HDC (HarmonyOS NEXT), WebDriverAgent (iOS)
• Model serving: vLLM or SGLang (self-hosted) or BigModel/ModelScope API
• Input: Screenshot + task description → Output: structured action commands

Installation

Prerequisites

• Python 3.10+
• ADB installed and in PATH (Android) or HDC (HarmonyOS) or WebDriverAgent (iOS)
• Android device with Developer Mode + USB Debugging enabled
• ADB Keyboard APK installed on Android device (for text input)

Install the framework

git clone https://github.com/zai-org/Open-AutoGLM.git
cd Open-AutoGLM
pip install -r requirements.txt
pip install -e .

Verify ADB connection

# Android
adb devices
# Expected: emulator-5554   device

# HarmonyOS NEXT
hdc list targets
# Expected: 7001005458323933328a01bce01c2500

Model Deployment Options

Option A: Third-party API (Recommended for quick start)

BigModel (ZhipuAI)

export BIGMODEL_API_KEY="your-bigmodel-api-key"
python main.py \
  --base-url https://open.bigmodel.cn/api/paas/v4 \
  --model "autoglm-phone" \
  --apikey $BIGMODEL_API_KEY \
  "打开美团搜索附近的火锅店"

ModelScope

export MODELSCOPE_API_KEY="your-modelscope-api-key"
python main.py \
  --base-url https://api-inference.modelscope.cn/v1 \
  --model "ZhipuAI/AutoGLM-Phone-9B" \
  --apikey $MODELSCOPE_API_KEY \
  "open Meituan and find nearby hotpot"

Option B: Self-hosted with vLLM

# Install vLLM (or use official Docker: docker pull vllm/vllm-openai:v0.12.0)
pip install vllm

# Start model server (strictly follow these parameters)
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

Option C: Self-hosted with SGLang

# Install SGLang or use: docker pull lmsysorg/sglang:v0.5.6.post1
# Inside container: pip install nvidia-cudnn-cu12==9.16.0.29

python3 -m sglang.launch_server \
  --model-path zai-org/AutoGLM-Phone-9B \
  --served-model-name autoglm-phone-9b \
  --context-length 25480 \
  --mm-enable-dp-encoder \
  --mm-process-config '{"image":{"max_pixels":5000000}}' \
  --port 8000

Verify deployment

python scripts/check_deployment_cn.py \
  --base-url http://localhost:8000/v1 \
  --model autoglm-phone-9b

Expected output includes a <think>...</think> block followed by <answer>do(action="Launch", app="..."). If the chain-of-thought is very short or garbled, the model deployment has failed.

Running the Agent

Basic CLI usage

# Android device (default)
python main.py \
  --base-url http://localhost:8000/v1 \
  --model autoglm-phone-9b \
  "打开小红书搜索美食"

# HarmonyOS device
python main.py \
  --base-url http://localhost:8000/v1 \
  --model autoglm-phone-9b \
  --device-type hdc \
  "打开设置查看WiFi"

# Multilingual model for English apps
python main.py \
  --base-url http://localhost:8000/v1 \
  --model autoglm-phone-9b-multilingual \
  "Open Instagram and search for travel photos"

Key CLI parameters

Python API Usage

Basic agent invocation

from phone_agent import PhoneAgent
from phone_agent.config import AgentConfig

config = AgentConfig(
    base_url="http://localhost:8000/v1",
    model="autoglm-phone-9b",
    device_type="adb",  # or "hdc" for HarmonyOS
)

agent = PhoneAgent(config)

# Run a task
result = agent.run("打开淘宝搜索蓝牙耳机")
print(result)

Custom task with device selection

from phone_agent import PhoneAgent
from phone_agent.config import AgentConfig
import os

config = AgentConfig(
    base_url=os.environ["MODEL_BASE_URL"],
    model=os.environ["MODEL_NAME"],
    apikey=os.environ.get("MODEL_API_KEY"),
    device_type="adb",
    device_id="emulator-5554",  # specific device
)

agent = PhoneAgent(config)

# Task with sensitive operation confirmation
result = agent.run(
    "在京东购买最便宜的蓝牙耳机",
    confirm_sensitive=True  # prompt user before purchase actions
)

Direct model API call (for testing/integration)

import openai
import base64
import os
from pathlib import Path

client = openai.OpenAI(
    base_url=os.environ["MODEL_BASE_URL"],
    api_key=os.environ.get("MODEL_API_KEY", "dummy"),
)

# Load screenshot
screenshot_path = "screenshot.png"
with open(screenshot_path, "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

response = client.chat.completions.create(
    model="autoglm-phone-9b",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,{image_b64}"},
                },
                {
                    "type": "text",
                    "text": "Task: 搜索附近的咖啡店\nCurrent step: Navigate to search",
                },
            ],
        }
    ],
)

print(response.choices[0].message.content)
# Output format: <think>...</think>\n<answer>do(action="...", ...)

Parsing model action output

import re

def parse_action(model_output: str) -> dict:
    """Parse AutoGLM model output into structured action."""
    # Extract answer block
    answer_match = re.search(r'<answer>(.*?)(?:</answer>|$)', model_output, re.DOTALL)
    if not answer_match:
        return {"action": "unknown"}
    
    answer = answer_match.group(1).strip()
    
    # Parse do() call
    # Format: do(action="ActionName", param1="value1", param2="value2")
    action_match = re.search(r'do\(action="([^"]+)"(.*?)\)', answer, re.DOTALL)
    if not action_match:
        return {"action": "unknown", "raw": answer}
    
    action_name = action_match.group(1)
    params_str = action_match.group(2)
    
    # Parse parameters
    params = {}
    for param_match in re.finditer(r'(\w+)="([^"]*)"', params_str):
        params[param_match.group(1)] = param_match.group(2)
    
    return {"action": action_name, **params}

# Example usage
output = '<think>需要启动京东</think>\n<answer>do(action="Launch", app="京东")'
action = parse_action(output)
# {"action": "Launch", "app": "京东"}

ADB Device Control Patterns

Common ADB operations used by the agent

import subprocess

def take_screenshot(device_id: str = None) -> bytes:
    """Capture current device screen."""
    cmd = ["adb"]
    if device_id:
        cmd.extend(["-s", device_id])
    cmd.extend(["exec-out", "screencap", "-p"])
    result = subprocess.run(cmd, capture_output=True)
    return result.stdout

def send_tap(x: int, y: int, device_id: str = None):
    """Tap at screen coordinates."""
    cmd = ["adb"]
    if device_id:
        cmd.extend(["-s", device_id])
    cmd.extend(["shell", "input", "tap", str(x), str(y)])
    subprocess.run(cmd)