01On-site: keep real-time analysis close to the camera
Install the processor on a compatible Linux or Unraid Docker host in your equipment room. It reads locally reachable KVM or RTSP sources and runs object detection and real-time rule evaluation there. Supported NVIDIA GPUs accelerate inference; compatibility depends on the GPU and container build.
Local analysis does not mean zero data leaves the site. The standard service still uses our public API for configuration, status and uploaded evidence. Strictly offline or fully private control-plane deployments require a separate assessment.
02Managed: use our compute without buying a dedicated server
We operate the analysis processor. Approved remote sources connect through a Tailscale private network, avoiding direct Internet exposure of your KVM. You retain the existing cameras and NVR.
In this mode, the selected video stream travels to our processor for analysis. It is not on-premises-only processing. Tailscale access grants, network reachability, available bandwidth and your consent determine what can be connected.
03Know exactly what is uploaded
The processor initiates authenticated HTTPS requests to the API. It receives configuration and submits matching events, evidence screenshots, selected historical keyframes and operational telemetry. Device credentials and processor tokens must still be protected.
Historical keyframes support search even before a rule exists; they can be uploaded without a matching rule. They are sampled and deduplicated, not an upload of every 2 FPS frame. The server may generate semantic search indexes from retained images. Cloud storage therefore contains images as well as event metadata.
Rule events and historical keyframes have no hourly count or aggregate image-byte quota. Keyframe uploads require no paid plan. Image size, API request rates, sampling, local cache and retention remain bounded; semantic search still depends on your plan. This is not unlimited video storage or guaranteed delivery.
04Private connections, clearly defined
Tailscale uses WireGuard to encrypt traffic between participating endpoints, including traffic relayed through DERP. This protects the KVM-to-processor path when that path is routed through Tailscale. Ordinary RTSP over a local LAN is not automatically encrypted just because the processor has Tailscale installed.
Our public website and API use HTTPS and are published through Cloudflare Tunnel. The origin initiates the tunnel, so customer-side port forwarding and a publicly exposed origin are not required. The website and API themselves remain public endpoints protected by authentication; “no exposed origin” does not mean “no public API”.
05Start with a processor, then add your views
Activate a processor, add reachable video sources and choose which views participate in analysis. Define objects, confidence, regions and rules, then choose notifications. Test with a known event and check the evidence before relying on an alert.
Keep Docker and GPU drivers maintained, restrict Tailscale access to required devices, use unique credentials and enable account 2FA. Only connect cameras you are authorized to process.
Common questions
Does the camera need to support H.264?
For direct RTSP in the current build, yes. With KVM HDMI capture, the NVR performs its own camera decoding and supplies the displayed picture; supported camera codecs then depend on the NVR rather than our RTSP decoder.
Is every AI task performed inside my building?
Real-time object detection and rule checks run on the assigned processor. That processor is local in on-site mode and hosted in managed mode. Semantic search indexing of uploaded evidence can run on the data server.
Does this replace the NVR recording?
No. Keep the NVR as the continuous recording source. We retain selected evidence according to your account policy, not a full continuous backup.
Tailscale · WireGuard encryption ↗Cloudflare Tunnel ↗Privacy Policy →