Lavague

LaVague uses a world model approach to web automation — an LLM generates and executes browser actions based on natural language instructions and the current page state. Primarily research-oriented but increasingly production-capable for structured web tasks. Python-first, works with Selenium or Playwright as the browser backend. Good for teams building web navigation agents who want more than raw Playwright but less infrastructure than hosted solutions. Confidence is docs-derived.

Python SDK: pip install lavague. WorldModel-based architecture: world model interprets page context and instructions, action engine executes browser operations. Supports multi-step navigation with state management. Browser backends: Selenium or Playwright (configurable). Custom actions extensible via plugin architecture.

LLM key required (OpenAI/Anthropic via env variables). No LaVague-specific API key — open-source library. Selenium/Playwright driver credentials handled by standard browser setup. Open-source access model with no hosted auth complexity.

Browser automation reliability tracks Selenium/Playwright stability. LLM-generated action quality depends on model capability and page complexity. Complex pages (heavy JavaScript, iframes, shadow DOM) can reduce action accuracy. Python exceptions surfaced on automation failures. Active GitHub with regular releases.

docs.lavague.ai covers installation, world model concept, action engine, multi-step navigation, and integration guides. Getting started: pip install lavague, configure LLM key, first web navigation in under 15 minutes. Open-source MIT (GitHub: lavague-ai/LaVague). Community via LaVague Discord.