FAQ

Which LLM does it use? None is bundled. Any MCP client brings its own model; with Claude Code that is the Claude model you are logged into.

Why not just use screenshots? Screenshots are expensive (vision agents used ~45× more input tokens than API agents in one benchmark) and give the model nothing to act on directly. Text views with refs are cheaper and more precise. screenshot exists for the cases where layout matters.

Does it get past Cloudflare? It doesn't try to sneak past. It identifies itself (User-Agent token, Web Bot Auth signatures) and hands CAPTCHAs to a person. A logged-in session you provide already carries trust.

Is my data sent anywhere? The engine sends nothing by itself. What your agent sends to its model is exactly the text the tools return, after masking and scrubbing.

Why does --default-off exist? Default-deny is the safest start for a browser with real logins in it. With it (or a host app that owns policy), every site is Off until you allow it.

Does it read cross-origin iframes? Not yet. They are listed as iframe "host" cross-origin but their contents are not expanded. Closed shadow roots, canvas-only apps and rows of a virtualized list that are not rendered are also not visible to it.

Does it record and replay workflows? The session trace records every call with its tokens and timing, but replaying a workflow without a model is not available.

Does the human-like cursor help get past bot detection? No, and it is not meant to. It exists so people can follow the agent and so demo videos look natural. The engine still identifies itself and hands CAPTCHAs to a person.