Releases

Releases

OpenNVR ships in the open, most weeks. This page is the short version, newest first; every item links to the work behind it.

Updated 16 November 2026 · All releases on GitHub · Full changelog

On main, in the next release

since v0.1.12

v0.1.12 has just shipped; new work on main will be listed here as it lands.

    Open for discussion: Agentic Workflows

    A Workflows page where you build automations from steps — a trigger, rules, an AI step that looks or investigates, a person's approval, actions — on your own hardware. The design is public and we want your review.Read the RFC and comment.

    v0.1.12 · Setup, updates and apps without hand edits

    16 November 2026

    A new site installs releases by itself, first-time setup confirms the authenticator before it finishes, and the installers offer one-click app install, search by meaning and voice search, none of which needs the Camera Agent any more. A recorder's channels are added in one go, and ./start.sh uninstall gives a clean slate. App SDK 0.13.0.

    • Updates install by themselves by defaultA new site's update policy was off, so it never learned a release existed. The default is now automatic: a signed stable release is installed at night after the site's own delay, backed up first and rolled back if the new version is not healthy. Settings turns this down to notify-only or off, and a site that saved a policy keeps it.
    • A recorder's channels, added in one goConnecting to a discovered NVR or DVR lists its channels, with which are already cameras here, and adds the ticked ones together, each with its own main and sub stream.
    • First-time setup confirms the authenticatorSetup has a third step: a code from the app confirms the authenticator and the recovery codes are shown once and acknowledged. Closing the tab before scanning used to lock the operator out of the account they had just created.
    • One-click app install and voice search, offered by the installerThe installer asks whether apps may be installed and removed with one click, and whether to switch on voice search. Voice search and visit descriptions no longer need the Camera Agent installed.
    • ./start.sh uninstall: a clean slateRemoves this install from the machine, every app and its data included, after showing the plan and asking for a typed confirmation. Recordings and backups can be kept, and nothing of any other stack on the machine is touched.
    • A search result opens pausedClicking a result opens the recording paused on the instant, so the moment is not gone before the operator finds the subject.
    • The storage floor, and the "Degraded" it causedThe automatic rotation floor is 5% of the recordings disk, between 5 and 20 GB, which leaves room to pull an update. A full disk that rotates no longer shows as Degraded.
    • For app developers: a live description without a model call, and one guideAn app can ask what is on a camera now and get the description core already has. The rules for building an efficient app are one guide, carried into every scaffolded app and checked by opennvr-app validate --strict.
    Full v0.1.12 release notes →

    v0.1.11 · The Qualcomm NPU release

    14 November 2026

    On the qualcomm flavour the detector, the captions, the camera agent's language model, search, speech to text and the plate reader all run on the Hexagon NPU, and a fresh install on a Dragonwing board starts with one setting. Search gains voice, synonyms and typo reading; the workflow engine runs workflows; USB webcams are cameras. App SDK 0.12.0.

    • Captions on the Qualcomm NPUOn the qualcomm flavour the captioner is Gemma 4 E2B through LiteRT-LM, compiled for the Hexagon: 1.6 to 1.8 s a photo where the smallest CPU vision model took 16 s, the image encoder on the NPU too. The same bundle serves the camera agent, 52 of 52 on its live eval at 0.6 to 3 s a turn.#803
    • Speech to text on the Hexagon, by defaultWhisper Small (w8a16) as a QNN adapter: 361 ms an utterance against 774 ms for Parakeet on the ARM cores, equal accuracy on clean speech, and a no-speech gate. STT_ADAPTER=auto picks it on Qualcomm boards and Parakeet elsewhere.#839 · #845
    • Search on the NPUSigLIP 2 compiled for the Hexagon embeds a frame in 6 to 8 ms and a query in 4 to 6 ms; the caption side of search gets its own sentence encoder and a labelled 33-case eval that took mean reciprocal rank from 0.46 to 0.71.#855 · #864
    • The NPU on the dashboardThe NPU and the GPU sit beside CPU, memory and disk, and the tile says when requests are queued behind each other rather than capping at 100%. The ANPR contention measurements that led there are in the bench folder.#813 · #879
    • Plate reading on a Qualcomm boxCompiled graphs for the plate reader survive a recreate, two tuning knobs reach core, and the docs say what the plate stack actually is.#878
    • USB webcams, served by MediaMTXOn Linux a webcam is passed into the MediaMTX container and served with a main and a sub stream, like an IP camera; Docker Desktop on macOS and Windows uses the host agent. Nothing to install on a Linux host.#889
    • Voice searchA microphone button in the search box. The clip is transcribed on the box by the same speech-to-text adapter the camera agent uses; nothing leaves it, and the browser's own cloud speech recognition is not used.
    • Add Camera finds ONVIF cameras by multicastOne probe and every camera on the segment answers within seconds, with its name and model, and no subnet to configure. The address sweep still covers routed subnets.
    • The workflow engine runs workflowsAn app's alert starts the workflows it triggers; a run walks its steps (If, For each, an AI step, actions) and records each one. An Ask a person step waits for an approval. There is an API for both.#773
    • OpenNVR MirrorThe camera agent serves a kiosk page for a show floor: live video with detection boxes, a spoken sentence about the scene every few seconds, and a card per person. All on the box; nothing stored.
    • Detection boxes on by defaultEvery live picture shows its detection boxes without a setting.
    Full v0.1.11 release notes →

    v0.1.10 · A benchmark per device

    31 October 2026

    A small release on the same day: the detector benchmark gives one row per device so an NPU host can be measured with one command, and the fallback metric now really shows when an accelerator fell back to CPU.

    • Per-device detector benchmarkpython -m detect_pipeline.bench --latency N times the Tier-0 detector on whole frames per provider, names the provider ONNX Runtime really ran on, and exits non-zero when the accelerator fell back to CPU.
    • tier0_detector_backend shows a fallbackThe requested provider is kept when ONNX Runtime drops one or a session is rebuilt on CPU, so requested != effective fires in exactly the cases it exists for.
    Full v0.1.10 release notes →

    v0.1.9 · The Hexagon NPU, and OpenNVR updates itself

    31 October 2026

    OpenNVR Accelerate: hardware flavours of the same images, picked by one setting. On an Arduino Ventuno Q (Qualcomm Dragonwing IQ-8275) the Tier-0 detector and the ONNX adapters run on the Hexagon NPU: 8.6 ms an inference, sixteen cameras at 2 fps where the CPU managed four. Intel iGPU and NPU through OpenVINO. And a signed release manifest so ./start.sh update can switch, check and roll back.

    • Qualcomm Hexagon NPU (the -qualcomm flavour)OPENNVR_ACCEL=auto picks it on an arm64 host with the Hexagon compute DSP. The detector runs on the HTP through the ONNX Runtime QNN plugin with its compiled graph cached; the five ONNX adapters (detection, pose, faces, plates, packages) follow on their -qualcomm images, two to eight times faster end to end than on the ARM cores. Measured on boards from Indus Technologies.#734 · #735 · #736
    • Intel iGPU and NPU (the -intel flavour)The detector on the NPU or iGPU through OpenVINO and VA-API decode on the iGPU; each adapter on the device it is proven on, CPU otherwise, and every fallback logged and exported.
    • OpenNVR updates itself, if you let itEach release ships a signed manifest naming its commit, image digests and database revisions; ./start.sh update checks it, backs up, switches, verifies health and rolls back on failure.
    • Cameras page: bulk actions, grid view, pause recordingTick cameras and act on all of them; a grid view of stills with stream and recording state; recording can be paused per camera or in bulk for a chosen time.
    • Tour editor redesignOne row for a tour's name and state, settings in a dialog, each page says how many tiles have cameras, and a running wall edits the page it is on.
    Full v0.1.9 release notes →

    v0.1.8 · Better defaults, each one measured

    16 October 2026

    The agent's model, its speech to text, the always-on detector and the embedding model all change, and every change is a number with the harness that produced it beside it. The catalog says what hardware an app needs and whether this box has it. Adapter release 0.1.8 ships alongside, same version number.

    • The agent thinks with Granite 4granite4:micro is the default: 44/60 on the Workflows Investigate step where qwen3:1.7b passed 21/60, because it calls the tools it is given before deciding; 45/45 on the chat cases. Small talk no longer reaches the model with tools.#653 · #652
    • It hears with ParakeetSpeech to text is Parakeet TDT 0.6B v3: on the agent's own phrasings, a third of whisper's errors at a fifth of the time on four cores. Whisper stays for translation or Hindi; Piper stays for the voice.#658
    • It sees with YOLO26n, and remembers with SigLIP 2The detector is YOLO26n, 17% faster on CPU and more accurate with the same decode. The embedding model behind search is SigLIP 2, and a model change behind the adapter now re-embeds the store on its own.#654 · #659
    • Describe what to watch forOpen-vocabulary detection is in the catalog: a box and a score for any phrase, baked weights, no network, on demand.#646
    • The catalog says what hardware it needsEvery task carries its bar as numbers; an app inherits its tasks' bar and each card says fits, fits tightly or short for the box it is on. Advisory, never a gate.#661
    • Measured, in the repoThe SDK's Look step and the Investigate eval cases, by Emmanuel Nwanguma; harnesses for VQA, speech, voice and retrieval beside the agent, with their results.#651 · #649 · #655 · #660
    Full v0.1.8 release notes →

    v0.1.7 · Security release

    13 October 2026

    Five API surfaces checked that a caller was signed in but not which cameras they may see. Any account could read every camera's AI detections and captions, read a camera's configuration with its credentials, run inference on any camera, list every camera's recordings, and see the compliance pages for the whole site. Upgrade every install that has more than one user.

    • Per-camera authorization on five more surfacesAI detection results, camera configuration, inference, the AI recording listings and the compliance pages now follow the same rule as every other read surface: view permission to read, manage to write or delete, and 404 for anything outside your cameras. Reported by Artur Chlebicki; four related findings were found while triaging.#635
    • App SDK: model clients and shared tool definitionsThe first piece of the Agentic Workflows AI step: one chat() over Ollama and OpenAI-compatible servers with schema-constrained output, and the read tools' definitions shared between the camera agent and core.#636 · #637
    • Smaller fixesUser management names the field that failed validation, and Guard Scan Compliance installs from the App Catalog again.#634 · #629
    Full v0.1.7 release notes →

    v0.1.6 · Search your footage in a sentence

    11 October 2026

    147 merges. Footage becomes searchable in plain words, catalog apps become products with pages of their own, and both SDKs reach PyPI: App SDK 0.7.0 and Adapter SDK 1.3.0.

    • Search your footage in a sentenceA Search page over every recorded visit: "red truck at the dock yesterday" becomes filters you can edit, a person is a name you pick, and every result can show where the object went next, across cameras.#490 · #513 · #559 · #572
    • Visits are described, and search understands descriptionsCaptions, colours, vehicle types and plates become searchable. With the CLIP adapter, a search for "truck" finds the visit described as "a lorry"; a question that needs a description not written yet says "not yet", not "no".#503 · #510 · #546 · #552 · #566 · #593
    • Apps became productsLine Crossing, Loitering, Intrusion, Abandoned Object, Package Delivery, Gate Controller, Alert Notifier and Smart Doorbell each have a page of their own, and Guard Scan Compliance is new. Each app is configured on its page and runs only on the cameras picked for it.#476 · #477 · #486 · #488 · #489 · #492 · #498 · #499 · #574
    • A camera's AI follows its appsSelect cameras for an app and the skills it needs switch on for them; disable it and they switch off unless another app still needs them.#482 · #573
    • Home Assistant through MQTT discoveryOpenNVR's cameras, sensors and alerts appear in Home Assistant's MQTT integration, on a scoped API token, with integration secrets encrypted at rest. The native integration, with live video and a media browser, is now installable from HACS (released 14 October).#491 · #578
    • A faster, measured camera agentSimple questions are answered before the model is asked, CPU turns are quicker, ten small models were measured on real tool use, and the installer picks from your hardware.#594 · #595 · #609 · #616 · #617 · #619
    • Detections on the live view, one box per objectBoxes with label and score over the live video, and a slow detector drops frames instead of the camera session.#464 · #601 · #603
    • A redesigned dashboard and interface, in English or FrenchHealth, live cameras, alarms and trends on one rearrangeable screen; every view works in the light theme; and the interface speaks French.#474 · #580 · #582 · #588 · #604
    • Vehicles and playbackPlate scanning is fast, accurate or auto per camera; the plate reader moves to a newer model that also reports the region; and recordings play in Firefox and with G.711 audio.#600 · #605 · #610
    • Security fixesThe media server's control interface now requires a token, a forged X-Forwarded-For header no longer sets the client address, and an app can attach findings only to visits on its own cameras. Upgrading is recommended.#471 · #491 · #542
    • Both SDKs on PyPIApp SDK 0.7.0 brings the one-function App facade and opennvr-app dev; Adapter SDK 1.3.0 brings the Adapter front door and the opennvr-adapter command line.#462
    Full v0.1.6 release notes →

    v0.1.5 · The app platform grows up

    10 September 2026

    The largest release since 0.1.0, with a security sweep: upgrading is recommended for every install.

    • Every app has its own identityIts own credential, its own place on the event bus with permissions from its manifest, and network access only through an enforced egress proxy to the hosts it declared.
    • Signed apps, built from sourceCatalog images are signed and verified at install, and any app repository can build from source through a shared workflow.
    • Apps can be soldA licence-key entitlement and a verified badge in the catalog. OpenNVR takes no cut.
    • A real App CatalogSearch, categories, sorting, a Featured row, and detail pages for apps you have not installed.
    • Security fixesIncluding cross-camera event disclosure on the live WebSocket, a blind SSRF on camera create, and webhooks with no target restriction. Per-camera access control now covers every surface.
    Full v0.1.5 release notes →

    v0.1.4 · Detection that respects your CPU

    26 August 2026

    99 merges. The detection pipeline gets a CPU budget it keeps, and the camera agent becomes a front end rather than a demo.

    • Per-camera detectionChoose what each camera watches for, with an editor that suggests and explains assignments.
    • Decode settings that cut CPU without losing detectionsAdaptive decoding reads keyframes while a scene is quiet and goes full rate on activity.
    • A measured detector swapAn RF-DETR detector head and an evaluation harness, so changing detector is measured rather than hoped.
    • Fixes worth upgrading forThe default decoder setting stopped detection from starting, hardware decoding was configured but never used, and visit history was lost for cameras with named ids.
    Full v0.1.4 release notes →

    v0.1.3 · Memory

    18 August 2026

    Events and their evidence photos are kept, so the system can be asked about the past.

    • An event and evidence storeEvery finished visit is saved with its best photo, survives restarts, and can be searched.
    • The agent looks backIt searches history, describes a past event from its photo, and narrates a time window from the recordings.
    • Plate reads on vehicle visitsWith a guard against repeated OCR and an attributed audit trail.
    • Recordings you cannot lose by accidentDelete is separate from deactivate with a recycle bin, and footage keeps its camera identity through a database reset.
    Full v0.1.3 release notes →

    Earlier releases