Unblink

Unblink is an open-source AI camera monitoring application that separates capture, routing, and vision inference into relay, node, and worker components. It forwards camera streams from private networks to a public relay and broadcasts frame events to whitelisted AI workers which return detections, summaries, and alerts stored for later search.

Key Features

• Federated architecture with three roles: Relay (public router/multiplexer), Node (private network proxy), and Worker (AI vision processors).
• WebSocket-based relay protocol with multiplexed logical bridges for nodes, workers, and browser clients.
• Supports common camera transports including RTSP and MJPEG and exposes an HTTP API for browser clients.
• Frame extraction and a computer-vision event bus that emits FrameEvent and FrameBatchEvent messages for workers to consume.
• Workers download frame binaries and emit bidirectional results (detections, summaries, alerts) that are stored and searchable.
• CBOR used for protocol message encoding and a minimal protocol design focused on separation of concerns and efficient forwarding of raw bytes.
• Configurable node that creates a local config file and can be installed via the Go toolchain; runtime and deployment are intended to be self-hosted and extensible.

Use Cases

• Add AI-powered detection, summarization, and alerting to home, office, or NVR camera fleets while keeping cameras on local networks.
• Run custom vision models (including VLMs) as federated workers to perform specialized inference tasks and push structured results back to the system.
• Build privacy-conscious monitoring workflows where raw bytes are proxied and inference can be performed on dedicated or self-hosted worker infrastructure.

Limitations and Considerations

• The project provides components (relay, node, worker) but expects operators to run or host the relay or use an available public relay; no official hosted SaaS is bundled with the repository.
• Worker inference for modern vision-language models can require GPU resources and model-serving infrastructure; deploying large VLMs may need specialized hardware and serving stacks.
• The system forwards raw camera bytes through the node/relay and uses CBOR for messages; integrators should plan for network, TLS, and operational setup when exposing relays and APIs.

Unblink is focused on modular, federated camera monitoring and AI inference, enabling teams to integrate custom vision workers and store searchable event results. It is suitable for projects that require on‑prem camera access with cloud‑reachable coordination and extensible worker-based inference.