Four ways teams put it to work.
Each of these runs on endpoints that ship today — sessions, CDP, perform-web-task,
the native MCP server, and the /content snapshot. Where a capability is roadmap, it's
labeled inline, not implied.
Agentic workflows via MCP
Your MCP client (Claude, Cursor) needs to actually browse the web.
Point it at our native MCP server; browser_navigate / snapshot / click / type / screenshot become tools backed by real sessions.
Multi-turn lead-gen: Claude compiled a ByteDesk ICP list from GitHub. We pointed Claude at the built-in MCP server (cmd/mcp)
and asked it to profile ByteDesk's market. It called browser_navigate and
browser_snapshot across GitHub topic pages, read each result, and compiled a live
Ideal Customer Profile lead list — every row taken off a real page.
Autonomous web tasks
You want an agent to fill forms, research, and navigate multi-step flows.
perform-web-task takes a prompt (and optional output schema), drives the browser through a snapshot→model→action loop, and returns a structured result.
Live perform-web-task result + streamed steps.
Data extraction / scraping
You need structured data from pages at scale.
Create sessions, drive them over CDP, and pull an accessibility snapshot from /content. Route through a proxy and turn on extra stealth when a site pushes back.
Sessions + CDP + /content snapshot endpoint.
QA / E2E testing
You want browser tests against real cloud sessions.
Point Playwright or Puppeteer at a session’s cdp_url and run your existing suite — the connection is a drop-in.
Unmodified connectOverCDP drop-in proof.
Star it. Run it. Ship agents on it.
Open-source and free to self-host. Managed cloud is coming.