Capability · Framework — agents
LaVague
LaVague pioneered the large-action-model pattern: a world model plans the next step, a retriever finds the most relevant DOM chunk, and an action engine generates Selenium or Playwright code that actually runs in the browser. It's popular in agent research and for building bespoke vertical browser agents.
Framework facts
- Category
- agents
- Language
- Python
- License
- Apache-2.0
- Repository
- https://github.com/lavague-ai/LaVague
Install
pip install lavagueQuickstart
from lavague.core import ActionEngine, WorldModel
from lavague.core.agents import WebAgent
from lavague.drivers.selenium import SeleniumDriver
driver = SeleniumDriver(headless=False)
agent = WebAgent(WorldModel(), ActionEngine(driver))
agent.get('https://huggingface.co/docs')
agent.run('Find the Transformers quickstart page.')Alternatives
- browser-use — lighter Playwright-first library
- Skyvern — self-hosted platform with vision
- Stagehand — TypeScript DOM-first agent
- AgentQL — query-language over Playwright
Frequently asked questions
What makes LaVague different?
It explicitly separates planning (world model) from grounding (action engine) and emits inspectable Selenium/Playwright code rather than a black-box trace. That makes it easier to audit and replay.
Does LaVague need vision models?
No. The default flow is text-only, using DOM retrieval. Vision is optional and can be layered in for harder sites.
Sources
- LaVague — GitHub — accessed 2026-04-20
- LaVague — docs — accessed 2026-04-20