Novita Agent Sandbox と BrowserUse の統合

pip install browser-use

export NOVITA_API_KEY="<Your Novita AI API Key>"
export LLM_API_KEY="<Your Novita AI API Key>"
export LLM_BASE_URL=https://api.novita.ai/openai
export LLM_MODEL="<Your LLM Model ID>"

import asyncio
import base64
import os
import time
import dotenv
dotenv.load_dotenv(override=True)
from browser_use import Agent, BrowserSession
from browser_use.llm import ChatOpenAI
from novita_sandbox.core import Sandbox

async def screenshot(agent: Agent):
  # Screenshot function
  print("Taking screenshot...")
  page = await agent.browser_session.get_current_page()
  screenshot_bytes = await page.screenshot(format='png')
  # screenshot method returns the binary data of the image, we should save it as a PNG file
  screenshots_dir = os.path.join(".", "screenshots")
  os.makedirs(screenshots_dir, exist_ok=True)
  screenshot_path = os.path.join(screenshots_dir, f"{time.time()}.png")
  if isinstance(screenshot_bytes, str):
    screenshot_data = base64.b64decode(screenshot_bytes)
  else:
    screenshot_data = screenshot_bytes
  with open(screenshot_path, "wb") as f:
    f.write(screenshot_data)
  print(f"Screenshot saved to {screenshot_path}")

async def main():
    # Create Novita sandbox instance
    sandbox = Sandbox.create(
        timeout=600,  # Timeout in seconds
        template="browser-chromium",  # This template contains chromium browser and exposes port 9223 for remote connection
    )
    try:
        # Get Chrome debug port address from sandbox
        host = sandbox.get_host(9223) # Get sandbox port 9223 address
        cdp_url = f"https://{host}"
        print(f"Chrome Debug Protocol URL: {cdp_url}")
        # Create BrowserUse session
        browser_session = BrowserSession(cdp_url=cdp_url)
        await browser_session.start()
        print("BrowserUse session created successfully")
        # Create AI Agent
        agent = Agent(
            task="Go to hackernews and find the top 3 stories",
            llm=ChatOpenAI(
                api_key=os.getenv("LLM_API_KEY"),
                base_url=os.getenv("LLM_BASE_URL"),
                model=os.getenv("LLM_MODEL"),
                temperature=1
            ),
            browser_session=browser_session,
        )
        # Run Agent task
        print("Starting Agent task execution...")
        await agent.run(
            on_step_end=screenshot, # Take screenshot after each step
        )
        # Close browser session
        await browser_session.kill()
        print("Task execution completed")
    finally:
        # Clean up sandbox resources
        sandbox.kill()
        print("Sandbox resources cleaned up")
        exit

if __name__ == "__main__":
    asyncio.run(main())