Stagehand V2: FULLY FREE Browser Use AI Agent
Source: YouTube — WorldofAI, published 2025-06-13 Tool covered: Stagehand (Browserbase)
Summary
Walkthrough of Stagehand v2 — Browserbase’s open-source AI browser-automation framework built on Playwright. v2 ships self-healing automations (retry-on-DOM-drift) and full agent integration via MCP (works inside Cline and Cursor as a sub-agent, or as a standalone Python lib). Three primary CLI verbs — navigate, act, extract, observe — make natural-language browser tasks ergonomic. Real demos: scraping top-10 YouTube videos in ~3 minutes, and an auto-job-application example at ~10¢/hour.
Key claims
- Built on Playwright — inherits the most robust browser-automation primitives
- Self-healing: when the DOM changes, Stagehand retries with a different selector strategy instead of breaking
- MCP-native — installs into Cline via the MCP marketplace; Cursor works the same way
- Browserbase backend required (serverless browser runtime; free sandbox available)
- Cost: ~10¢/hour for complex automations (Browserbase fees)
Model evaluation
Stagehand published model benchmarks alongside v2:
- Fastest + cheapest: DeepSeek R1 distilled Qwen 32B
- Most accurate (multimodal): Gemini 2.0 Flash
The fact that the fastest and cheapest slot goes to a distilled DeepSeek matches the wiki’s broader DeepSeek R1 thread.
Why it matters
Stagehand fills the browser-automation slot in the wiki’s open-source agent landscape — until now browser-use was the only entry, and it’s a Python lib without a clean MCP story. Stagehand is the polished MCP-native option that drops into Cline / Cursor without leaving the IDE.
Pairs with the Computer Agent slot in CodeLLM (this batch) — both are “the agent that controls the browser” but Stagehand is open-source and OSS-host-agnostic.