The Zero-Hardware Local AI Architecture
Not every builder has a 24GB RTX 4090 on their desk. By combining a local orchestrator (n8n, Agent Zero, or LangGraph running on a laptop or cheap VPS) with free cloud GPU runtimes (Google Colab T4 or Kaggle Dual-T4), you get 15GB–30GB of VRAM for $0.00.
Why Outbound WebSocket Relays Beat Standard Tunnels
Standard trycloudflare.com quick tunnels frequently trigger Cloudflare 1016 / 1033 edge blocks when called from another Cloudflare Worker or automated webhook.
- Method 1 (Direct Quick Tunnel): Run
cloudflared tunnel --url http://127.0.0.1:8080inside Colab/Kaggle and paste the generatedhttps://*.trycloudflare.com/v1URL directly into your local n8n instance. - Method 2 (Durable Object Reverse WebSocket Relay): The Colab/Kaggle runtime dials outbound via standard HTTPS WebSocket to
/connecton your Cloudflare Worker (ollama-relay). Your agents always call the exact same permanent Worker URL (/v1/chat/completions), even when the underlying Colab session restarts!