Introduction

OpenNVR records your IP cameras and runs AI over the footage, on hardware you own. No cloud account, no per-camera licence, no vendor holding your video.

It is a normal NVR first — discovery, recording, retention, live view, playback, users and permissions — with three things a normal NVR does not have.

The three things

Cameras cannot reach the internet

Cameras are the weakest device on most networks: default credentials, firmware nobody updates, and a manufacturer’s cloud they phone home to. OpenNVR assumes that and puts them behind it.

On a dual-NIC host the cameras sit on their own subnet with no route out. The recorder reaches them; they reach nothing. Apps you install run on a third network with no route to the cameras or the internet, and any host an app wants to reach is a grant you make explicitly and can see in the audit log.

You choose the models, and they stay yours

Inference is decoupled from recording. A small always-on detector — Tier-0 — runs motion, region selection, detection and tracking on every camera, and anything heavier is an adapter: a container that advertises what it can do and answers when asked.

That means you can run YOLOv8 on a CPU offline, swap in a better plate reader next month, or point at a cloud provider if you decide the trade is worth it. Reaching a cloud model is off unless you turn it on, and audited when you do. Model weights are fingerprinted, so a model swapped underneath a deployment is detectable rather than silent.

It answers questions

The OpenNVR Agent turns the estate into something you can ask. What’s at the back gate right now? runs a real inference on a live frame and answers from what it saw — locally, with the tool calls and their timings visible, so you can tell the difference between an answer and a guess.

What you actually get

Recording	Continuous and event-based, per-camera retention, seek-to-time playback
Live video	WebRTC in the browser, multi-camera wall, substreams for low bandwidth
Detection	Always-on tracking on every camera with no adapter required
AI	Any conformant adapter — detection, pose, plates, faces, captions, speech
Apps	A one-click catalog, each app declaring what it needs and what it can reach
Governance	Per-camera permissions, MFA, an encrypted credential vault, your own TLS keys, audit trail
Integration	REST, WebRTC, ONVIF, RTSP, MQTT, webhooks, a NATS event bus

Everything runs in Docker. The same images run on a Raspberry Pi 5, a laptop or a rack server; the difference is how many cameras and how heavy a model.

Where to go next

To run it — Docker setup is the ten-minute path, then first-time setup for the admin account and MFA.

To understand it — Architecture overview shows where each piece sits, and Platform features is the full capability list.

To build on it — the developer corner covers both SDKs: one for apps, one for publishing a model. An app is nine lines, and the platform gives it detections, a settings form, zones the operator draws, and the alert inbox. The three references are generated from the code: the App SDK, the Adapter SDK, and the operator API.

To evaluate it — OpenNVR compared against Frigate, ZoneMinder, Shinobi and Verkada, and the §889 assessment for procurement.

The licence, briefly

The platform is AGPL-3.0. Both SDKs are Apache-2.0, so an app or adapter you write is yours under any licence you choose, closed included, and OpenNVR takes no fee on what you sell. The architecture is documented in a peer-reviewed paper.