shared browser control + cinematic recording studio for agents. agent and human co-pilot the same chromium — vision path is pixels, not DOM snapshots. MIT.
capabilities
- always live-visible — CDP screencast streams ~30fps to a WebSocket viewer; models ask `screen` and get pixels.
- shared session — human and agent operate the same chromium simultaneously. keyed by sessionId.
- cinematic recordings — declarative .tape scripts drive browser + recording together; effects are first-class verbs.
- vision built in — Ollama moondream/llava fallback for hosts without first-class image input.
- HITL native — the Ask verb pauses a tape, prompts the user through the viewer, fills a variable, resumes.
