Creativity · Agent Protocol

LaVague: Large Action Model Web Agent Framework

LaVague, open-sourced in 2024, is a web agent framework centered on a Large Action Model (LAM): a small, fine-tuned model that maps natural-language web instructions into concrete browser actions (Selenium or Playwright). LaVague combines a Retriever (find the right DOM context) with a LAM (emit an action), producing a transparent, auditable, lightweight alternative to frontier-model web agents.

Protocol facts

Sponsor
LaVague AI (open source)
Status
stable
Spec
https://github.com/lavague-ai/LaVague
Interop with
Selenium, Playwright, Hugging Face, LlamaIndex

Frequently asked questions

What is a Large Action Model?

A LAM is a model trained to emit actions (clicks, typing, URL visits) rather than freeform text. LaVague popularized the term for a fine-tuned small model that is fast and cheap but specialized to web-action prediction.

How does LaVague handle complex pages?

A Retriever first selects the minimal DOM slice relevant to the instruction; the LAM then emits an action scoped to that slice. This keeps context small and reliable on real-world pages.

Is it still active in 2026?

LaVague's open-source repo is still maintained and widely cited in research; it's used as a baseline and a teaching tool even where production deployments prefer frontier-model pipelines.

Sources

  1. LaVague GitHub — accessed 2026-04-20
  2. LaVague docs — accessed 2026-04-20