Apache-2.0 SDKs · your licence · no revenue share

Build on OpenNVR

Cameras, always-on detection, pluggable inference, an event bus, recordings, an operator alert inbox, users with per-camera permissions, and a catalog every deployment opens — all already running. You write the rule, the model or the workflow.

OpenNVR is a framework as much as a product: the lower layers are done, so an application never touches RTSP streams, frame cropping, best-frame selection or the detector. It receives events with evidence, calls the model it needs through one contract, and holds the business logic. Bring an expensive model and spend it only on the frames that matter. The App Catalog is what people have built this way.

Three ways in. Code: the two SDKs below, for an app or a model. No code to write: Agentic Workflows, a visual builder and a sentence that drafts the workflow, in design onRFC #620.Headless: an AI assistant that builds from the published skill definitions and the workflow file format, so a workflow made anywhere can be imported with an operator's approval.

Two SDKs, and which one you want depends on a single question.

opennvr-app-sdk

I want to ship an app

A rule, an analytic, a workflow, an integration. You get detections off the bus, a settings form generated from your parameters, zones the operator draws on a camera still, and alerts that land in their inbox — without writing any of it.

  • Nine lines is a complete, catalog-listed app
  • Declare a task; any adapter serving it satisfies you
  • Recordings, timeline search and evidence photos by API
uvx --from opennvr-app-sdk opennvr-app new driveway-watch
# or: pip install opennvr-app-sdk && opennvr-app new driveway-watch

opennvr-adapter-sdk

I want to publish a model

A detector, a captioner, an ASR model, anything that takes a frame or a clip and answers. Declare the tasks it serves and the SDK gives you the six contract endpoints, auth, metrics, the failure envelope and the streaming protocol.

  • A conformance kit you run before anyone else does
  • OpenAPI and AsyncAPI documents generated, not maintained
  • Weights fingerprinted, so tampering is detectable
uv add opennvr-adapter-sdk
# or: pip install opennvr-adapter-sdk
# then wrap your model in about thirty lines

Looking for something to build? Build this next lists the apps and adapters people keep asking for, and what each builds on.

The one idea that governs both

An app asks for a task, never for an adapter by name. An app declaring requires_tasks: [license_plate_recognition] is satisfied by any adapter advertising that task — so a better plate reader is a drop-in, and an app written today works with a model published next year.

It cuts both ways. Invent a task name nobody asks for and your adapter installs, reports healthy and receives no work — a failure that is completely silent. Both SDKs check this for you, and the catalog index validator fails CI over it.

the handshake

# the app says what it needs
requires_tasks: [object_detection]

# the adapter says what it serves
tasks_advertised: [object_detection]

# nothing else has to agree

Contribute to OpenNVR itself

The platform is built in the open, and help is welcome at every size: an interface fix, a test, a page of docs, a new adapter.

Reference

Your code stays yours

Both SDKs are Apache-2.0. Ship under any licence, closed included, even though the platform core is AGPL.

No fee, no cut

OpenNVR takes nothing on app or model sales. You get the licence hook; the platform stays out of the transaction.

Distribution included, never required

Every deployment opens the same catalog, and listing is a pull request against a public index. You never have to list: an app or adapter run from your own repository works exactly the same.