Stagehand

Stagehand occupies the middle ground between brittle Playwright scripts and unpredictable end-to-end browser agents: developers choose per step whether to write deterministic code or delegate to AI through four primitives, act for single actions, extract for schema-validated data, observe for discovering possible actions, and agent for full multi-step tasks. Automations self-heal when page layouts change, and a caching layer replays previously seen actions without further LLM inference, cutting cost and latency on repeated runs. Built by Browserbase on the Chrome DevTools Protocol, it ships first-class TypeScript and Python SDKs with alpha bindings for other languages, scaffolds new projects via npx create-browser-app, and runs against local browsers or Browserbase's hosted browser infrastructure with most major LLM providers. The framework is MIT licensed with about 23,700 GitHub stars. Teams use it for scraping, form automation, and QA in cases where pure selectors are too fragile and pure agents too unreliable.