What it does Unblink is a camera monitoring stack that pairs a local Go node with a cloud relay. The node sits on your network, pulls video from RTSP or MJPEG cameras via go2rtc, and forwards it to a central server. The server handles auth, config, and storage. You interact through a SolidJS frontend that lets you chat with your feeds and search recorded frames in plain English.

The interesting bit The VLM angle is what separates this from yet another NVR. Qwen3-VL analyzes frames for summarization, which means the search isn’t just timestamp or motion-based — it’s semantic. “Show me the delivery person in a red jacket” is the pitch, though how well that works in practice depends on the model and your hardware.

Key highlights

• Local node runs as a single Go binary with prebuilt releases for Linux and Windows (x86_64 and ARM64)
• Relay-node architecture keeps camera traffic off the public internet; only the control plane hits the cloud
• go2rtc handles RTSP/WebRTC translation, which is a solid, battle-tested choice
• Frontend stack is modern but conventional: SolidJS, TypeScript, Vite, Ark UI
• PostgreSQL with pgx for persistence; protobuf for internal comms

Caveats

• The README is thin on actual VLM performance: no benchmarks, no latency numbers, no hardware requirements for local inference
• “AI-powered” and “real-time” are claimed but undefined — unclear if Qwen3-VL runs locally or in the cloud, and what that costs
• One-click Render deploy is offered, but running your own relay server plus a local node plus camera hardware is still a multi-piece setup

Verdict Worth a look if you want semantic search over camera footage without building the VLM plumbing yourself. Skip it if you need a drop-in replacement for Blue Iris or want proof that the natural-language features work well before committing to the architecture.