How Velvt works, layer by layer.
The collection agent is open-source. Everything above it is sealed on your machine until you choose otherwise.
L1 — Collection agent
Passive macOS background process. Logs active window titles, app names, and timestamps via the Accessibility API. Roughly one event per five seconds of active use. No screenshots, no keylogging, no content capture.
- Swift · Accessibility API · auditable source
- ~576 KB raw events per day
- Abstracted titles before any modeling
L2 — Behavioral modeling
Computes Focus Fragmentation, Switching Clusters, and Proportionality. Days 1–14 use rule-based thresholds; from day 15, deviations are measured against a rolling 14-day baseline (2× margin).
L3 — Insight generation
Lightweight model translates metrics into one neutral sentence. Deterministic syntactic fallback when offline. Voice of a data scientist, not a cheerleader.
L4 — Delivery
Menu bar app, native notifications, three timeline tracks. Maximum three alerts per day: one threshold, one inactivity, one evening summary at 8:00 PM.
or read our local-first and audit docs. Questions: hello@getvelvt.com