Using Browser Use as a CrewAI Agent Tool

Learn how to enable CrewAI agents to interact with web browsers using Browser Use

Browser Use is a popular new tool that allows AI to control web browsers and uses LLMs to generate the actions. While it can be used as its own type of AI Agent, I wanted to pair it with CrewAI to create a more powerful and flexible agent. As I've previously mentioned, I'm a big fan of using CrewAI to create agents that can perform a wide range of tasks. You can learn more about CrewAI in my previous post where I explore a full stack framework for building and deploying CrewAI agents.

CrewAI Tools

CrewAI has a powerful tooling system that allows you to create custom tools for your agents. My goal was to utilize Browser Use as a tool for my agents. Experimenting I found that it works best to wrapper Browser Use as a specific tool to achieve a certain outcome rather than a general purpose tool. One challenge I faced was that CrewAI tools don't natively support running async functions and Browser Use requires async functions to interact with the browser. Also my full stack framework already was using Python event loops which seemed to conflict with how I was trying to run the Browser Use tool.

CrewAI Browser Use Tool

To get around these issues, I created a custom CrewAI tool that uses the Browser Use API. Here's the code for the tool (note I am not a Python expert so this code is probably not perfect):

from pydantic import BaseModel, Field
from crewai.tools import BaseTool
from agents.privateagents.private.thc_agent.schemas import ProductList
from crew_agents.tools.websearch import use_browser
import asyncio
from concurrent.futures import ThreadPoolExecutor

class WebsiteInput(BaseModel):
    """Input schema for the Product Browser Tool"""
    website: str = Field(
        description="The website URL to search for products"
    )

class ProductBrowserTool(BaseTool):
    name: str = "Product Browser Tool"
    description: str = "Use a browser to search for products on the specified website"
    args_schema: type[BaseModel] = WebsiteInput

    async def _arun(self, website: str) -> str:
        result = await use_browser(
            f"""Go to {website} and search the entire site for products. (....more instructions....)""",
            ProductList,
            20
        )
        return result

    def run_async_code(self, website: str) -> str:
        """Run async code in a new event loop"""
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        result = loop.run_until_complete(self._arun(website))
        loop.close()
        return result

    def _run(self, website: str) -> str:
        with ThreadPoolExecutor() as executor:
            future = executor.submit(self.run_async_code, website)
            result = future.result()
             return result
    

Using the Tool

Now that we have the tool, we can use it in a CrewAI tasks, just pass it in the tools parameter.

Overall it seems to be working well for my use case. Hope this helps someone else looking to use Browser Use with CrewAI.