Skip to main content
This tutorial teaches you how to use GBOX as a provider in OS World to build and run agents that can interact with operating systems.

​
What is OS World?

OSWorld is a benchmark framework for evaluating multimodal agents on open-ended tasks in real computer environments. It supports multiple providers for running virtual environments, including VMware, VirtualBox, Docker, and AWS. By using GBOX as a provider, you can leverage cloud-native infrastructure without managing local virtual machines, making it easier to scale your agent evaluations and reduce setup complexity.

​
Architecture

The following diagram illustrates the architecture of OS World using GBOX as a provider:

​
Benefits of Using GBOX Provider

Using GBOX as a provider in OS World offers several advantages:

​
🚀 Cloud-Native Infrastructure

  • No need to set up and manage local virtual machines
  • Works seamlessly across different development environments

​
⚡ Easy Scaling & Parallelization

  • Run multiple environments in parallel without local resource constraints
  • Significantly reduce evaluation time through parallel execution

​
🔧 Simplified Setup

  • No need to check KVM support or install Docker Desktop
  • Works on any platform without virtualization requirements

​
🌐 Accessibility

  • Access your environments from anywhere
  • Consistent performance regardless of your local hardware

​
Prerequisites

Before getting started, make sure you have:
  • A GBOX account with an API key (Get your API key)
  • An OpenAI API key (or another compatible LLM provider)
  • Python 3.10 or higher installed
  • Git installed

​
Getting Started

​
Step 1: Clone the Repository

Clone the OS World provider repository: 
# Clone the OSWorld provider repository
git clone https://github.com/babelcloud/OSWorld-provider

# Change directory into the cloned repository
cd OSWorld-provider

# Optional: Create a Conda environment for OSWorld
# conda create -n osworld python=3.10
# conda activate osworld

# Install required dependencies
pip install -r requirements.txt

​
Step 2: Configure API Keys

Create a .env file in the repository root and add your GBOX API Key and OpenAI API Key:
.env
GBOX_API_KEY=your_gbox_api_key
OPENAI_API_KEY=your_openai_api_key
Note: You can obtain your GBOX API key from the API Key page. Make sure to keep your API keys secure and never commit them to version control.

​
Step 3: Run the Provider

Execute the following command to start the provider with GBOX: 
python run_multienv.py \
    --provider_name gbox \
    --model gpt-4o \
    --region us-east-1 \
    --max_steps 15 \
    --observation_type screenshot \
    --action_space pyautogui \
    --result_dir ./results_gbox \
    --num_envs 1 \
    --test_all_meta_path evaluation_examples/test_small.json
Command Parameters Explained:
  • --provider_name gbox: Use GBOX as the provider
  • --model gpt-4o: Specify the LLM model for the agent
  • --region us-east-1: GBOX region (adjust based on your preference)
  • --max_steps 15: Maximum number of steps the agent can take
  • --observation_type screenshot: Use screenshots for environment observation
  • --action_space pyautogui: Use PyAutoGUI for action execution
  • --result_dir ./results_gbox: Directory to save evaluation results
  • --num_envs 1: Number of parallel environments to run. Increasing this value can significantly improve evaluation efficiency by running multiple tasks concurrently
  • --test_all_meta_path: Path to the test configuration file

​
Step 4: Monitor Agent Execution

Once the agent starts running, you can monitor its progress in real-time through the VNC viewer. The agent will interact with the OS environment, performing tasks based on the evaluation configuration. VNC View
Tip: The default VNC password is osworld-public-evaluation. You can access the VNC viewer URL from the GBOX dashboard or API response.

​
Step 5: View Results

After the evaluation completes, you can find the results in the results_gbox directory. The results include:
  • Task execution logs
  • Screenshots of key actions
  • Performance metrics
  • Success/failure status for each task
You can now start building your own agents by modifying the test configuration files or creating custom evaluation scenarios.

​
Next Steps

  • Explore the OSWorld documentation to learn more about creating custom evaluation tasks
  • Check out the GBOX API reference for advanced configuration options
  • Experiment with different models and parameters to optimize agent performance
  • Scale up your evaluations by increasing the --num_envs parameter to run multiple environments in parallel