Three picks, then your first scan runs.
Defaults are good. You can override anything before you kick off the scan.
Pick any of these to skip the form and jump straight to a ready-to-run scan against a real bundled demo site.
The page the wizard references. This wizard plays a seeded demo replay — agents don't actually navigate this URL. To run a real scan against a live URL (including file:// paths to local HTML), use the Live multi-agent scan panel below.
The journeys agents will attempt. Pick one or many — the scan runs one agent fleet per journey. At least one required. (You can toggle these anytime, even after advancing steps.)
AgentGlass runs all three browser agents in parallel: Anthropic Claude Computer Use, OpenAI Operator (gpt-5.5), and Gemini Baseline (cheap deterministic + LLM probes). Set LIVE_AGENTS in the backend env to gate which ones actually fire.
Three browser agents (Claude Computer Use, OpenAI Operator, Gemini Baseline) run in parallel against your URL. Each one tries the canonical task, emits per-step traces, and contributes findings. An adversarial Claude verifier reviews the merged set and discards duplicates / false-positives.