XDA wrote up a build last week that paired Frigate with a local LLM through Home Assistant, then connected an 8B vision model to actually describe what the cameras were seeing. The result is the upgrade self-hosters have been chasing for years: motion alerts that say “a delivery driver dropped a package at the door at 11:42” instead of “motion detected at front camera.” The reason it works is that all three pieces are local, so the vision model is not gated by an API rate limit.
We tested seven desktop and self-hosted NVR apps that pull AI directly into the camera pipeline. The picks run on Windows, macOS, Linux, or a mix of all three, and pair with the popular local model servers like Ollama, vLLM, and LM Studio. We ranked by what they actually do with the video stream: object detection, facial recognition, descriptive alerts, and pairing with Home Assistant or HomeKit.
What to look for in an AI-powered NVR
Five things separate an NVR that runs object detection from one that turns alerts into something useful:
- True local inference. Cloud-based AI bills by the frame, which is why most Ring or Nest plans cap detection. Local detection on a GPU or Coral TPU is free past the hardware cost.
- RTSP-first ingestion. ONVIF discovery is convenient but RTSP is what the open NVRs speak. Camera support stops at this line.
- Hardware acceleration. NVIDIA NVENC, Intel QuickSync, or Coral USB Accelerator for object detection at speed. A CPU-only setup ages fast.
- Home Assistant integration. The killer feature of the XDA build is that the LLM ran through HA so it could pull motion events and describe them. The integration quality matters.
- Modular AI back-ends. Pluggable inference servers (CodeProject.AI, Frigate+) beats a black-box bundled model the vendor controls.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price | AI support |
|---|---|---|---|---|---|
| Frigate NVR | Object detection + HA integration | Linux, Docker, Proxmox | Open-source | Free | Coral TPU, ONNX, GPU |
| CodeProject.AI Server | Modular AI back-end | Windows, Linux, macOS, Docker | Open-source | Free | Multi-model |
| Blue Iris | Heavy commercial NVR | Windows | $69.95 one-time | $69.95 | DeepStack, CPAI plugin |
| AgentDVR | Cross-platform NVR | Windows, macOS, Linux | Free tier | $7.99/mo Premium | DeepStack, ONVIF |
| Viseron | Lightweight container NVR | Docker on Linux, macOS | Open-source | Free | Edge TPU, CPU, GPU |
| Shinobi CCTV | Pro features open-source | Windows, macOS, Linux | Open-source | Free | TF object detection |
| MotionEye | Raspberry Pi-class NVR | Linux, Docker | Open-source | Free | None bundled, plugin |
1. Frigate NVR, best for object detection + Home Assistant integration
Frigate NVR is the open-source NVR the self-hosting world settled on for one reason: the Home Assistant integration is first-class. Frigate runs object detection in real time using a Coral TPU, NVIDIA GPU, or modern Intel iGPU, publishes detection events over MQTT, and feeds Home Assistant a live thumbnail per camera. The XDA build that triggered this article connected the Frigate events to a local Llama 3 vision model running on the same box, then routed the descriptions back into HA notifications.
Where it falls short: Docker-first, which means the setup curve is steeper than Blue Iris if you have not run a container before. Windows is supported only through WSL or Docker Desktop, not native.
Pricing:
- Free: open-source, MIT
- Paid: Frigate+ subscription unlocks higher-quality detection models, $50/year, optional
Platforms: Linux native, Docker on Windows/macOS, Proxmox, Home Assistant OS add-on
Download: Frigate NVR
Bottom line: the NVR for the home lab that already runs Home Assistant. The local LLM pipeline runs on top of Frigate cleanly, and the upgrade path from object detection to descriptive alerts is documented.
2. CodeProject.AI Server, best modular AI back-end
CodeProject.AI Server is not an NVR. It is the AI inference server the other NVRs plug into. Drop CPAI on the same machine as Blue Iris, AgentDVR, or Shinobi, and it exposes object detection, face recognition, licence plate reading, and superresolution as HTTP endpoints the NVRs query. The model library is modular: install only the ones you need, swap them out without touching the NVR.
Where it falls short: on its own, CPAI does no recording. Pair with an NVR. The Windows installer is friendlier than the Linux setup, which still leans Docker.
Pricing:
- Free: open-source, modules included
- Paid: none
Platforms: Windows, Linux, macOS, Docker (NVIDIA, ROCm, Intel HW acceleration)
Download: CodeProject.AI Server
Bottom line: the back-end of choice if you want to layer AI onto an existing NVR. Pairs especially well with Blue Iris.
3. Blue Iris, best heavy commercial NVR
Blue Iris is the long-standing Windows-only commercial NVR that drives a lot of small-business installs. The 2026 build integrates CodeProject.AI as a first-party AI module, so object detection, face matching, and licence-plate reads are configured in the same UI as the camera profiles. Recording quality, motion zones, and PTZ camera support are the deepest on this list.
Where it falls short: Windows-only, $69.95 buys version 5 with one-year support, then $35 per major upgrade. Not cheap by self-hosted standards.
Pricing:
- Free: 15-day trial
- Paid: $34.95 lite (one camera) or $69.95 full, plus optional yearly support
Platforms: Windows 10 and 11
Download: Blue Iris
Bottom line: the right pick for a Windows-only home lab that wants every PTZ feature and is willing to pay once. CPAI handles the AI side.
4. AgentDVR, best cross-platform NVR
AgentDVR is the cross-platform answer to Blue Iris, written by the iSpyConnect team. The same NVR core runs on Windows, macOS, Linux, and Docker. AI integrations support DeepStack and CodeProject.AI for detection, and the web UI handles remote viewing without a separate app. The free tier covers most home use; Premium adds cloud recording, advanced motion, and longer event history.
Where it falls short: the AI integration is less polished than Frigate’s Home Assistant pipeline. Description alerts via local LLM need more glue code than Frigate’s MQTT events.
Pricing:
- Free: full feature set, 2 cameras, cloud features limited
- Paid: Premium from $7.99/mo for full cloud and unlimited cameras
Platforms: Windows, macOS, Linux, Docker, Raspberry Pi
Download: AgentDVR
Bottom line: the cross-platform NVR for users who want one tool that runs on their Mac or Linux machine.
5. Viseron, best lightweight container NVR
Viseron is the Docker-native NVR that targets the same Home Assistant audience as Frigate but trades polish for flexibility. The object detector supports Edge TPU, ONNX, and CPU back-ends. Configuration is a single YAML file. Stream recording happens on motion only, so disk usage stays low on a small server.
Where it falls short: smaller community than Frigate, fewer integrations, and the documentation lags behind features. Best for users comfortable reading source when stuck.
Pricing:
- Free: open-source, Apache 2.0
- Paid: none
Platforms: Docker on Linux and macOS
Download: Viseron on GitHub
Bottom line: the secondary pick if Frigate’s Docker layout does not fit. Stronger on resource efficiency than feature depth.
6. Shinobi CCTV, best for pro features without paying
Shinobi CCTV is the open-source NVR with pro polish, originally built for installers who wanted Blue Iris features without the licence. The TensorFlow object detection runs in the same process, and the dashboard exposes timeline scrubbing, multi-camera grids, and event filtering closer to a paid product.
Where it falls short: development pace varies, and the AI module is older than Frigate’s. The maintainer has been responsive but releases are not as frequent.
Pricing:
- Free: open-source, MIT
- Paid: optional support subscription from Shinobi.video
Platforms: Windows, macOS, Linux, Docker
Download: Shinobi CCTV
Bottom line: the right pick if you want a polished dashboard and a free licence. AI integrations are usable but lag Frigate.
7. MotionEye, best for Raspberry Pi-class hardware
MotionEye is the lightweight NVR that has been running on Raspberry Pi boards for a decade. It does not bundle modern AI, but it can pipe RTSP streams into Frigate or CodeProject.AI on a separate machine, then feed the alerts back into Home Assistant. The MotionEyeOS image makes a Pi 4 into a four-camera NVR with a few clicks.
Where it falls short: no native AI, the dashboard is functional rather than slick, and event search is rudimentary. It is meant to be paired with smarter tools.
Pricing:
- Free: open-source, GPL
- Paid: none
Platforms: Linux, Docker, MotionEyeOS image for Raspberry Pi
Download: MotionEye on GitHub
Bottom line: the lightweight ingestor for a Pi-class machine. Pair with Frigate or CodeProject.AI for the AI layer.
How to pick the right one
- If you already run Home Assistant and want the XDA-style descriptive-alert build, install Frigate NVR as the camera back-end and pair it with a local LLM through HA. Nothing else on this list is as well-documented for that flow.
- If you are Windows-only, Blue Iris plus CodeProject.AI Server is the most polished commercial-feeling stack. Pay once, learn the UI, done.
- If you want cross-platform and dislike Docker, AgentDVR is the answer.
- For a Mac home lab that wants the Frigate experience, AgentDVR runs natively. Frigate via Docker also works but the macOS Docker overhead is real.
- For low-resource Raspberry Pi setups, MotionEye ingests cameras and a second machine handles the AI.
- If you want the AI layer separately so you can swap the NVR later, install CodeProject.AI Server first. Most NVRs on this list integrate with it.
FAQ
What is Frigate and why does the AI-camera community keep recommending it?
Frigate is an open-source NVR built around real-time object detection on Coral TPUs or GPUs. The Home Assistant integration publishes detection events as MQTT messages, which makes it easy to wire into automations, dashboards, or, as in the XDA build, a local LLM that describes what was detected.
How do I add a local LLM to my security camera setup?
The reliable flow: Frigate detects an object and publishes the snapshot to MQTT. Home Assistant subscribes to the event and triggers a script that sends the snapshot to a vision-capable local LLM (Llama 3.2 Vision, MiniCPM, Qwen2-VL) hosted in Ollama or LM Studio. The LLM returns a description, HA pushes a notification with that text. The XDA piece walks through the full configuration.
What hardware do I need for local AI camera analysis?
The minimum useful setup is a Coral USB Accelerator ($60) plugged into a Pi 5 or small N100 box, which handles object detection for 4-6 streams. For descriptive alerts via vision LLM, a consumer GPU with 12GB VRAM (RTX 3060 or RTX 4060 Ti) runs an 8B vision model in real time. An older 8GB card also works with smaller models.
Is local AI better than Ring or Nest cloud detection?
For privacy, yes by definition: the video never leaves your network. For accuracy, the gap closed in 2025. A current Frigate+Coral setup with a fine-tuned model identifies people, vehicles, animals, and packages with comparable accuracy to a Nest Aware subscription, and the description quality from a local 8B vision model beats the canned templates the cloud services send.
What is the best free AI-powered home security camera app?
Frigate NVR for the NVR side, CodeProject.AI Server for the inference side. Both are open-source, mature, and integrate with Home Assistant. The combined stack costs nothing past the hardware.
