LaVague is an open-source framework for building AI web agents, built around the Large Action Model idea. The project has about 6,400 stars on GitHub, with its most recent activity in 2025.
LaVague splits a web agent into two parts. A World Model looks at the current page and the user’s goal and decides what should happen next, while an Action Engine compiles those decisions into Selenium or Playwright code that actually drives the browser โ code you can read and debug instead of a black box. Agents can run against headless browsers or drive your own Chrome session through the LaVague extension, which helps with sites that require a logged-in, real browser. A Gradio interface makes demos easy, and the companion LaVague QA tool turns Gherkin specifications into automated web tests, giving the framework a second life in quality assurance. LaVague grew out of research on Large Action Models โ models meant to act in interfaces rather than only chat about them. The framework abstracts model choice, so builders can pair the World Model with different LLMs and trade cost against reliability, and the browser code it generates can be saved and reused as ordinary automation scripts.
LaVague is free and open source under the Apache-2.0 license. The only running cost is the LLM API usage your agents generate.
Developers building browser agents for automation and research, and QA engineers interested in generating web tests from natural-language specifications.