browser-use Self-Hosted Agent

browser-use is an open-source browser-controlling agent – an LLM navigates a real Chrome instance, clicks, types, and extracts data. On our dedicated GPU hosting it is a legitimate production path for web automation, data gathering, and QA workflows.

Contents

• Stack
• Deployment
• Vision vs text-only
• Tips

Stack

• browser-use Python library
• Playwright + headless Chromium
• LLM via OpenAI-compatible API (vLLM)
• Optional: vision-capable LLM for screenshot-based navigation

Deployment

pip install browser-use playwright
playwright install chromium

from browser_use import Agent
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="llama-3.3-70b",
    openai_api_base="http://localhost:8000/v1",
    openai_api_key="not-needed",
)

agent = Agent(
    task="Find current GPU server pricing on gigagpu.com and return a CSV",
    llm=llm,
)
result = await agent.run()

Vision vs Text-Only

browser-use can run in two modes:

• Text-only: feeds the agent a structured DOM representation
• Vision: captures screenshots and feeds them to a vision-capable LLM

Vision mode works better on visually-dense sites but requires a VLM – Qwen VL 2 7B or Llama 3.2 Vision 11B. Text-only works on more sites and runs faster.

Tips

• Chromium needs a display (xvfb) on headless servers
• Rate-limit your agent – web automation is easy to take too far
• Respect robots.txt and site terms
• Persist cookies across runs for authenticated sites

See Open Interpreter.