LaVague Tutorial: Automate Web Tasks with Natural Language Instructions

Getting Started with LaVague in 5 Minutes

In this tutorial, you'll build a working AI Web Agent that navigates Hugging Face's documentation and finds specific pages — all by describing what you want in plain English.

Prerequisites

• Python 3.9+
• An OpenAI API key ($5 credit is plenty to start)
• Chrome or Firefox browser (for Selenium)

Step 1: Install LaVague

pip install lavague

This installs the core framework along with Selenium driver support.

Step 2: Set Your API Key

export OPENAI_API_KEY="sk-your-key-here"

LaVague uses GPT-4o by default for both the World Model and Action Engine. You can configure alternative LLM backends — see the customization docs.

Step 3: Write Your First Agent

Create a file agent.py:

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

selenium_driver = SeleniumDriver(headless=True)
world_model = WorldModel()
action_engine = ActionEngine(selenium_driver)
agent = WebAgent(world_model, action_engine)

agent.get("https://huggingface.co/docs")
agent.run("Go on the quicktour of PEFT")

Step 4: Run It

python agent.py

Watch as LaVague navigates through the Hugging Face docs, finds the PEFT library page, and clicks through to the quick-tour section — all without you writing a single CSS selector or XPath expression.

Step 5: Add the Interactive Gradio Demo

LaVague includes a built-in Gradio interface for interactively testing your agents:

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

Customizing Your Agent

LaVague supports extensive customization:

Advanced: Try LaVague QA

For QA engineers, LaVague QA extends the framework to convert Gherkin feature files into automated browser tests. This bridges the gap between natural-language test specs and executable test suites — a game changer for teams practicing behavior-driven development.

Cost Considerations

Each agent step makes LLM calls. A typical 5-step task on a simple page costs approximately $0.01-0.03. For complex multi-page workflows, use the built-in token counter to estimate costs before running at scale.