What it does

Gabber is a self-hosted engine and visual editor for building real-time AI applications that ingest live media—microphone audio, camera feeds, screen capture—and process them through composable node graphs. It runs as four services (Next.js frontend, editor backend, execution engine, and a thin repository server) and connects to LiveKit for WebRTC media transport.

The interesting bit

The “pad” system is the wiring: typed ports on each node can operate in either Property mode (holding a value) or Stateless mode (streaming), which lets you mix event-driven and stateful logic in the same graph. Add a state machine on top and you get branching behavior—like an AI security guard that switches from idle to alert when motion exceeds a threshold.

Key highlights

• Visual graph editor with typed node connections (pads) and subgraph composition
• Built-in nodes for transcription, LLM inference (local via llama.cpp/vLLM or remote), emotion analysis, and media I/O
• State machines with parameter-driven transitions for multi-phase app logic
• Secrets stored outside graph definitions (.secret file), so graphs are shareable without credential leakage
• SDKs available for TypeScript, React, and Python; Unity, mobile, and Flutter marked “Coming Soon”
• Licensed under fair-code (Sustainable Use License), same model as n8n; examples and SDKs are Apache 2.0

Caveats

• Several SDKs (Unity, iOS, Android, React Native, Flutter) are listed as “Coming Soon” with no timeline
• Local LLM setup on macOS requires building llama.cpp from source; the brew package reportedly lacks Metal GPU support
• The repository service stores all apps and subgraphs in a local .gabber directory—no remote persistence or versioning mentioned

Verdict

Worth a look if you’re building voice/video AI agents and want a visual, self-hosted alternative to managed platforms. Skip it if you need mature mobile SDKs or a battle-tested deployment story beyond Docker Compose on localhost.