Usage

Minimal example

This example navigates to the HuggingFace docs and follows the PEFT quick-tour link -- the canonical LaVague demo from the official quick-tour guide.

from lavague.core import WorldModel, ActionEngine
from lavague.core.agents import WebAgent
from lavague.drivers.selenium import SeleniumDriver

# Open a Selenium-backed browser (headless=False shows the browser window)
driver = SeleniumDriver(headless=False)

# Build the three core components
world_model = WorldModel()
action_engine = ActionEngine(driver)

# Compose them into a WebAgent
agent = WebAgent(world_model, action_engine)

# Navigate to the starting URL
agent.get("https://huggingface.co/docs")

# Run the objective -- the agent loops until done or it gives up
agent.run("Go on the quicktour of PEFT")

---

Interactive demo mode

To launch the Gradio interface where you can step through each action and approve or reject it before it executes:

agent.demo("Go on the quicktour of PEFT")

This opens a local web UI at http://127.0.0.1:7860 where you can watch the agent reason, see what action it plans, and either confirm or stop it.

---

Environment variables

Variable	Required	Purpose
OPENAI_API_KEY	Yes (default LLM)	Powers the WorldModel and ActionEngine
LAVAGUE_TELEMETRY	No	Set to NONE to opt out of usage data collection

---

Notes

• headless=False is required when the target site has bot detection, CAPTCHAs, or login requirements -- these cannot be handled in a headless session without additional setup.
• The agent operates a real browser session. Do not point it at live production accounts without careful review of the objective.
• LaVague does not bundle its own LLM; it delegates to the configured provider (OpenAI by default). Costs accrue at the provider's token rates.