​
🌐 Browser Use Integration

Browser Use is a popular open-source browser agent built on LangChain. We’ve created a Lucidic-integrated fork so you can track and debug your browser agent workflows with full session replay, step timelines, event logging, and prompt evaluation — right out of the box.

​
🚀 What We’ve Added

In our fork of the Browser Use agent:

  • Lucidic calls like lai.create_step() and lai.end_step() are automatically inserted into the agent logic
  • LLM and tool interactions are tracked as Events
  • Support for Mass Sims, Prompt DB, and Rubric evaluations is built in
  • Works with LangChain handler for full call graph coverage

​
🧭 How to Use It

​
1. Clone the Lucidic Fork

git clone https://github.com/lucidic-ai/browser-use-lai.git
cd browser-use-lai

📘 This is a fork of the official repo with added observability.

​
2. Install Locally

pip install . ."[dev]"

This installs browseruse as a local package

​
3. Use the LangChain Handler

Since Browser Use is built on LangChain, you still need to attach the Lucidic handler manually.

In your script:

import lucidicai as lai
from browser_use.agent import BrowserAgent

# Initialize Lucidic
lai.init(
    session_name="browser_run",
    agent_id="your-agent-id",
    providers=["langchain"]
)

# Create agent and attach LangChain handler
agent = BrowserAgent()
handler = lai.LucidicLangchainHandler()
handler.attach_to_llms(agent)

# Run as usual
agent.run("Find the capital of Japan and report its population.")

​
4. Example Script

You can run full-scale testing of your browser agents using our example script on GitHub. It supports multiple models, prompt versions, and tasks — all fully tracked with Lucidic.

Below are a few key snippets:

​
🔁 Test Multiple Models

Loop through LLMs like GPT-4 and Gemini using a shared browser agent:

llm_configs = [
  {"name": "gpt-4.1-nano", "provider": "openai", "api_key": openai_key},
  {"name": "gemini-2.0-flash", "provider": "google", "api_key": gemini_key},
]

for llm in llm_configs:
    agent = Agent(llm=load_llm(llm), task="Search Hello world on google", browser=shared_browser)
    await agent.run()

​
🧠 Test Different Prompt Versions

Dynamically inject different versions of prompts (production, development, etc.) into your agents:

prompt_labels = ["production", "development"]

for label in prompt_labels:
    content = lai.get_prompt("System Prompt", label=label)
    system_message = SystemMessage(content=content.format(max_actions=4))

​
🧪 Run Mass Simulations

Track each run under a different session name and mass sim ID:

lai.init(
  task="Search Hello world on google",
  session_name=f"{prompt_label} | {llm_name}",
  mass_sim_id="a89c61f9-1a16-4a68-85f3-31cd2eca8abc"
)