Private by design. Provable by architecture.
What you capture from your feeds can be as personal as a diary. Security is Scroll Coach's first priority, above speed and above cost, and this page explains what that means in practice, without marketing fog.
The architecture
Isolation the database enforces
Every table that holds your data carries row level security with owner only policies. Your rows are invisible to every other account, enforced by the database engine itself. We verified it the blunt way: a second account sees zero rows.
AI access is read only
Any AI assistant you connect gets read only tools, rate limited and re-authenticated on every request. There is no write tool to abuse: this is our standing defense against prompt injection, built into the product's shape.
Media is ephemeral
When a video needs analysis, the media is processed and then deleted the instant analysis finishes, including on failure paths. Your corpus keeps the understanding: transcript, analysis, a thumbnail, and the link back. We are not a file locker.
Embeddings stay home
The numeric search index for your corpus is computed inside our own database environment. Your text is never sent to an external embeddings service. That was a deliberate, locked decision.
Delete means delete
Deleting is a true hard delete, on your device and in the cloud. Deleting everything revokes AI access tokens first, so nothing can read your corpus while it disappears. Export it all as JSON or Markdown first, any time.
No scraping, structurally
Capture happens on your own device, in your own signed in session, only when you initiate it. Platform passwords are never stored on our servers; sign ins never leave your phone. There is no crawler to turn evil.
What AI does touch your content.
Building the analysis is AI work, and we would rather tell you exactly how than hide behind vague words. Anthropic models write the analysis and power chat. Google Gemini watches videos. Content goes to them only to produce your result, and our providers do not use API content to train their models.
Analytics are content free: no autocapture, no titles, no URLs, only fixed event names like "a capture finished". Crash reports carry no personal identifiers.
Data is encrypted in transit everywhere. We do not claim end to end encryption, because that would be false: the service has to read content to analyze it. We say precisely what we do, and we do it.
The full details live in the Privacy Policy and Terms of Service.
The short list that matters.
A corpus you can trust with the personal stuff.
That is the whole point. Join the waitlist and see it firsthand.