ghostpilot streams a live chromium session while a vision model observes and acts alongside you. not a headless scraper — a shared workspace where human and machine collaborate in real time.
the vision layer runs locally via ollama. your screenshots never leave your machine. secret redaction is built in.
architecture
- browser layer — puppeteer-core with direct cdp screencast.
- vision layer — local ollama models, never cloud.
- control layer — ~40 rest endpoints for programmatic control.
- recording layer — capture to gif, mp4, or webm with zoom regions.
