Early access — free while in beta

Your cameras. Your AI.
Your data.

envyr is the cloud for the open-source sentry-x3 edge NVR. Detection and scene understanding run on your hardware; only the moments that matter sync to the cloud. Continuous footage never leaves your site.

Get early access See how it works ↓
No credit card · The edge is MIT-licensed · Works on a $50 board
LIVErear · beach path1.4s to live
person 91%
🔍
EVENTS · SYNCED TO YOUR POCKET
person · rearwalking along the path toward the beach
notify
wombat · leftgrazing at the edge of the dune grass
logged
white van · rightparked beside the access track
logged
How it works

The AI lives at the edge.
The cloud just carries the news.

A tiny edge box records continuously and runs YOLO on its NPU, gated by the cameras' own motion detection. A vision-language model describes what it sees. Only events — thumbnails, clips, descriptions — sync up.

1
📷

Record & detect locally

Continuous recording to your own disk. Camera motion wakes the NPU; YOLO classifies; a VLM writes a one-line description of the scene with movement and direction.

2
☁️

Sync only what matters

Events, thumbnails and short clips travel over the open envyr-sync protocol — mTLS device identity, outbound-only, offline-buffered. The 24/7 footage stays home.

3
📱

Review anywhere

Web dashboard and mobile apps with push that respects you: one alert per visit, not per frame. Tap into sub-2-second live view when you need eyes now.

your siteevents only →envyr cloud
Everything a tier-one NVR has

…without handing over your footage

🧠

Edge AI detection

YOLO on the edge box's NPU — people, vehicles, animals. Tuned for the wildlife big-brand cams throw away.

🔍

VLM descriptions

A vision-language model writes what happened: "a person in a white shirt walking toward the gate." Direction and all.

💬

Natural-language search

"Red car last Tuesday" finds it. Descriptions make the whole timeline conversational.

Instant live view

WebRTC straight off the camera substream — warm streams answer in under two seconds, anywhere.

🔔

Push that matters

One visit, one alert. Session dedup and per-day budgets — never thirty pings for one possum.

🛰️

Fleet management

Enroll a site with a claim code. Push camera config remotely. Watch health across every site you run.

💾

Smart retention

Recent footage in full-res; AI-flagged moments kept; the rest tiers down gracefully on your own disk.

🔓

No lock-in

The edge and sync protocol are MIT. Point the same box at your own server tomorrow — we have to earn you.

Privacy by architecture

Not a policy. A design constraint.

☁️ Typical cloud camera

  • 24/7 footage streams to someone else's datacenter
  • AI runs on their servers, on your private video
  • Subscription lapses → your cameras go dumb
  • Breach radius: everything you ever recorded

  • Continuous video stays on your hardware, full stop
  • Detection + scene AI run at the edge, on your NPU
  • Cloud down? The NVR keeps recording and detecting
  • Cloud holds events only — mTLS devices, tenant-isolated by database row-level security, signed expiring media links
Bytes of continuous footage uploaded: zero.
<2sto live video, warm-stream WebRTC
100%of AI inference at the edge
1push per visit — session-deduped
Open core

The edge is open source.
Forever.

sentry-x3 — recorder, detectors, VLM pipeline, timeline viewer — is MIT-licensed. envyr is the optional hosted backend, speaking the same open protocol anyone can implement.

  • Self-host everything on your LAN, free
  • Or enroll in envyr for apps, push and multi-site in one minute
  • Leave any time — your edge, your recordings, your call
★ sentry-x3 on GitHub
# run the open-source edge make build && ./bin/viewer -config config.yaml # …then add the cloud when you want it ./bin/enroll -code WMBT-4421 -endpoint envyr.io ✓ enrolled "beach-house" — 3 cameras syncing

Be first through the gate

Early access is open — free during beta, shaped by the people using it. Bring a $50 board and your cameras.

Request early access
or self-host the MIT edge today — no account needed