OpenAGI Lux + Kernel Browser Integration

Run OpenAGI's Lux computer-use models against serverless browsers powered by Kernel.

Demo

📹 Watch the demo video - Shows the Lux agent navigating to agiopen.org using a Kernel serverless browser.

What is OpenAGI?

OpenAGI is an AI research organization building foundation models for computer use. Their Lux model is a vision-language model specifically designed to control computers by:

• Analyzing screenshots to understand the current UI state
• Deciding on the next action (click, type, scroll, etc.)
• Executing actions in a screenshot-action loop until the task is complete

What is Kernel?

Kernel provides Browsers-as-a-Service for AI agents and browser automation. Key features:

• Serverless Browsers: Instantly launch browsers without managing infrastructure
• Computer Controls API: Native OS-level mouse, keyboard, and screenshot controls
• Stealth Mode: Built-in anti-detection for reliable web automation
• Video Replays: Record browser sessions as MP4 videos
• Scalability: Run hundreds of concurrent browser sessions

Model Selection

Choose the right model for your task:

Usage Approaches

How It Works

This integration connects OpenAGI's Lux model to Kernel's serverless browsers using custom providers:

1. KernelScreenshotProvider: Captures screenshots using Kernel's Computer Controls API
2. KernelActionHandler: Translates Lux actions (click, type, scroll) to Kernel commands
3. KernelBrowserSession: Manages browser lifecycle with automatic video recording

AsyncDefaultAgent Example

Use AsyncDefaultAgent for high-level, tasks with either the lux-actor-1 model for immediate tasks or lux-thinker-1 for more complex, longer running tasks:

async with KernelBrowserSession(
    record_replay=True,
    replay_output_path=replay_output,
) as session:
    # Create the screenshot provider and action handler
    provider = KernelScreenshotProvider(session)
    handler = KernelActionHandler(session)

    agent = AsyncDefaultAgent(
        api_key=os.getenv("OAGI_API_KEY"),
        max_steps=20,
        model="lux-actor-1"  # or "lux-thinker-1" for complex planning tasks
    )

    # Execute the task
    print(f"\nExecuting task: {instruction}\n")
    success = await agent.execute(
        instruction=instruction,
        action_handler=handler,
        image_provider=provider,
    )

TaskerAgent Example

Use TaskerAgent for structured workflows with predefined steps:

async with KernelBrowserSession(
    record_replay=True,
    replay_output_path=replay_output,
) as session:
    provider = KernelScreenshotProvider(session)
    handler = KernelActionHandler(session)

    tasker_agent = TaskerAgent(
        api_key=os.getenv("OAGI_API_KEY"),
        base_url=os.getenv("OAGI_BASE_URL", "https://api.agiopen.org"),
    )

    tasker_agent.set_task(
        task="Navigate to the 'What is Computer Use' section of the OAGI homepage.",
        todos=[
            "Go to https://agiopen.org.",
            "Make sure to press enter to start the navigation.",
            "Click on the 'What is Computer Use?' button."
        ]
    )

    result = await tasker_agent.execute(
        instruction="",
        action_handler=handler,
        image_provider=provider
    )

Setup

1. Install dependencies

uv pip install kernel oagi python-dotenv Pillow

2. Configure environment variables

Create a .env file with your API keys:

KERNEL_API_KEY=your_kernel_api_key
OAGI_API_KEY=your_openagi_api_key

Get your API keys:

• Kernel: dashboard.onkernel.com
• OpenAGI: developer.agiopen.org

3. Run the agent

For AsyncDefaultAgent (supports lux-actor-1 or lux-thinker-1):

python agent_default.py

For TaskerAgent (structured workflows with todos):

python agent_tasker.py

The agent will:

1. Launch a serverless browser via Kernel
2. Start recording a video replay
3. Execute the task using Lux's vision-action loop
4. Save the replay as video_agent_replay.mp4
5. Clean up the browser session

Project Structure

├── agent_default.py       # AsyncDefaultAgent example (lux-actor-1/lux-thinker-1)
├── agent_tasker.py        # TaskerAgent example with todos (lux-actor-1)
├── kernel_session.py      # Browser lifecycle & replay management
├── kernel_provider.py     # Screenshot provider using Kernel API
├── kernel_handler.py      # Action handler with key translation
├── pyproject.toml         # Project dependencies
└── video_agent_replay.mp4       # Recorded demo video

License

MIT