Concept · config-only · coming soon

Canary Vision Pro. Sees in the dark, listens for trouble.

The flagship sensor for the one spot in your fleet that actually needs it — starlight low-light imaging and on-device sound classification, wired into the same witness kernel as every $10 Canary. Not new firmware: a new tier, built on hardware that already exists.

Built on the Seeed reCamera Pro (2 GB) · SKU 100092895 · ~$299.90

What it is

A standalone AI camera, not an ESP32 peripheral

The reCamera Pro runs its own Linux OS and does its own on-device inference. SecuraCV's job is the same as it is for Frigate: accept its detections at the trust boundary, strip everything but a coarse claim, and seal it.

SoC

Rockchip RV1126B

3 TOPS NPU — vision, sound classification, and optional VLM/LLM inference, entirely on-device.

Sensor

8 MP starlight

Low-light imaging that resolves detail cheap PIR/RF sensors can only guess at.

Training

SenseCraft AI

1-click image/audio classification — no ML expertise needed to add or retrain a class.

Mount

PoE or USB-C

Magnetic mount / ¼″-20 tripod thread — a universal standard, bridged onto the catalog's own wall ecosystem.

How it plugs in

Configured, not flashed — five steps, zero new Rust

Point its built-in flow editor's HTTP action node at the existing webhook adapter, and it's a witness source. It inherits rate limiting, auth, TLS, and the kernel's zone/confidence/coarsening rules automatically.

reCamera Pro classClaim kindNotes
personLargeObjectBoundaryCrossingMatches the Frigate adapter's mapping.
car / vehicleLargeObjectBoundaryCrossingPair with an after-hours schedule to route to VehiclePresenceAfterHours.
package / dog / cat / bicycleSmallObjectBoundaryCrossing 
glass break / alarm / raised voiceAcousticImpulseInZoneSound classification only — never a waveform.
person in restricted zonePresenceInRestrictedZoneScoped by the flow's zone, not a face/re-id match.
For makers: the actual webhook payload

Point the reCamera Pro's HTTP request node here — the request path names the zone/sensor, the body is the shape the adapter already parses:

POST http://adapter-host:8800/sensors/loading_dock/camera
{"confidence": 0.91}

Add matching [[adapter.route]] entries to adapter_host.toml, set an auth_token or hmac_secret, and add TLS if the camera isn't on a trusted LAN. Full grammar in webhook.rs.

What it deliberately doesn't do

It can caption a scene. We don't route that.

No free text, ever

The payload grammar has no caption field

A caption literally cannot cross into a Claim — enforced in the sensor adapter contract, not just policy.

Classifier label only

Same discipline as every other sensor

Configure the on-device flow to emit a label + score — the same rule already used for Frigate's object labels.

VLM stays local

A convenience, never a witness input

An operator can pull up the camera's own Web UI during an active incident. That view never feeds the sealed log.

Where it fits

A $300 flagship, not a swap-in for a $10 sensor

Good fit

  • Multi-tenant stairwells, parking structures, loading docks after dark
  • One higher-value chokepoint in a small business — back door, register, equipment yard
  • Warehouse safety zones already running RS485/Modbus
  • A trusted anchor node in a mesh of cheaper sensors

Skip it here

  • A single-family porch — a $10 XIAO + PIR already covers it
  • Bedside / wellbeing monitoring — that's Canary Sense's camera-free mmWave lane

Nothing to buy yet. Something real to join.

No countdown clock, no fake backer count — just an open thread where interest becomes the next batch. Tell us where you'd put one.

Everything above is a concept, honestly labeled. The integration works today against the existing webhook adapter if you own the hardware; there is no SecuraCV-branded kit, bundle, or price yet. In the spirit of the product, a concept never pretends to ship.