CORTEX

Ambient HUD for NVIDIA DGX Spark clusters.

Your AI cluster deserves a better status light than nvidia-smi.

CORTEX is a full-screen, arc-reactor-style dashboard that sits on your rack monitor and shows what your cluster is actually doing — live GPU temps/util/power from every node, vLLM tokens/sec, running/waiting queues, CUDA kernel activity, agent tool calls. It shifts visual states between dormant → listen → thinking → hot → gold → fault so the room can tell at a glance whether the model is idle, generating, thermal-throttling, or broken.

Built for the 52Pi rack screen + DGX Spark combo. Works on any Linux box with a browser.

Features

• Live per-node GPU telemetry via nvidia-smi (local) + SSH (remote nodes)
• vLLM inference metrics polled from /metrics (tok/s, prompt/s, running, waiting, gen totals)
• State machine with visual modes:
  • dormant — standing by (blue ambient)
  • listen — incoming message (green pulse)
  • thinking — vLLM generating (cyan cascade + CRT terminal overlay)
  • hot — any GPU ≥ 80°C (amber)
  • gold — owner-keyword detected in inbound (gold shimmer)
  • fault — vLLM unreachable 3+ seconds (red alarm, flashing)
• Agent gateway log tail (optional) — emits tool, voice, msgIn, msgOut, reload events into the event stream
• Fully keyboard-demoable — press T/G/H/L/R for each mode, space for shockwave
• Zero config needed to boot — sensible DGX Spark defaults; env-overridable
• Pure static frontend — Node + ws backend, Three.js + Canvas2D + SVG frontend, no build step

Quick start

git clone <this-repo> cortex && cd cortex
npm install           # just `ws`
node server.js        # http+ws on :8090

Open http://localhost:8090 on your rack display (or use Firefox kiosk).

Configuration

Copy .env.example → .env (or export vars in your shell). All keys are optional; defaults show DGX Spark-friendly labels (NVIDIA DGX CLUSTER, 2× GB10 · 256GB UNIFIED, etc.).

Minimal single-node setup: just run node server.js. It polls local nvidia-smi + local vLLM on 127.0.0.1:8000.

Two-node cluster:

export NODE_2_SSH=user@10.0.0.2
export NODE_2_NAME=NODE-02
node server.js

With identity:

export AGENT_NAME=JARVIS
export OWNER_NAME=TONY
export OWNER_KEYWORDS=tony,stark,boss

Requirements

• Node 18+ (native fetch, WebSocketServer from ws@8)
• nvidia-smi on PATH (local node)
• SSH key auth to remote nodes if using NODE_2_SSH
• vLLM serving /metrics on a known endpoint (default 127.0.0.1:8000)

Running on the rack display

Spark 2 autologin + Firefox kiosk mode pointing at Spark 1's CORTEX server is the intended setup. Example systemd user unit and kiosk wrapper are left as an exercise — see your distro's kiosk docs.

Keyboard demo

While focused on the page:

Key	Mode
T	thinking (12s, 35 tps)
G	gold (12s)
H	hot (12s)
L	listen (6s)
R	reset to dormant
space	fire shockwave ring

Or use URL hashes: http://localhost:8090/#thinking, #gold, #hot, #listen, #dormant.

Architecture

           ┌─── nvidia-smi (local)
           │
 server.js ┼─── ssh nvidia-smi (NODE_2_SSH)   ─┐
(Node+ws)  │                                    ├─> state.gpus
           ├─── http GET vLLM /metrics (1Hz)  ──┼─> state.vllm (+ fault tracker)
           │                                    │
           └─── tail gateway-*.log (optional) ──┴─> state.events

           ws:/ws  ─────────────────────────────> browser
                                                   ├─ index.html (boot + corners)
                                                   ├─ cortex.js  (mode machine + orb)
                                                   ├─ terminal.js (CRT overlay)
                                                   └─ rain/constellation/css

License

MIT. Do whatever.

Disclaimer

Not affiliated with NVIDIA. DGX, Spark, GB10, and related marks belong to NVIDIA Corp.