Stagehand
Open-source AI browser-automation framework from Browserbase. Built on Playwright; v2 (June 2025) added self-healing retry logic and full MCP-native integration so it works as a sub-agent inside Cline and Cursor without leaving the IDE. Three primary verbs — navigate, act, extract, observe — make natural-language browser tasks ergonomic.
- Vendor: Browserbase
- Repo: github.com/browserbase/stagehand
- Built on: Playwright
- Backend: Browserbase serverless browser runtime (free sandbox available; ~10¢/hour for paid)
What’s new in v2
- Self-healing: when the DOM changes, Stagehand retries with alternative selector strategies instead of breaking
- MCP-native: drop-in install via Cline’s MCP marketplace; Cursor works the same way
- Standalone Python lib still supported
Model recommendations (from Stagehand’s own benchmarks)
| Slot | Model | Why |
|---|---|---|
| Fastest + cheapest | DeepSeek R1 distilled Qwen 32B | best speed/cost on Stagehand’s eval |
| Most accurate | Gemini 2.0 Flash | best multimodal grounding |
Why it matters
Stagehand fills the browser-automation slot in the wiki’s open-source agent landscape. Compared to browser-use (Python-only, no MCP story), Stagehand is the polished MCP-native option. Sibling to the Computer Agent inside CodeLLM (this batch) — same use case, but Stagehand is OSS and host-agnostic.
Sources
- Stagehand V2 FULLY FREE Browser Use AI Agent — WorldofAI walkthrough; YouTube scraping + auto-job-application demos