DROPLET SYSTEM

The droplet (129.212.180.252) is the public CUDA-MCP² broker. It proxies rtx:// jobs down a reverse tunnel to the Victus RTX 3050 — the edge that owns the silicon. Keys stay local; nothing GPU is public.

// for a new user — bring it live

The system is self-healing, but a fresh machine starts cold. Three moves:

1 · droplet (GPU-less broker)

ssh root@129.212.180.252
pip install -r requirements.txt   # vps_node deps
# run as a service so it survives droplet reboots:
cat > /etc/systemd/system/aevps.service <<'EOF'
[Unit]
Description=ae VPS node (CUDA-MCP² broker + PDS)
After=network-online.target
[Service]
ExecStart=/usr/bin/python3 /opt/aevps/vps_node.py --host 0.0.0.0 --port 3000
Restart=always
[Install]
WantedBy=multi-user.target
EOF
systemctl enable --now aevps

2 · Victus (the RTX leaf)

cd C:/æ/agents
bash rtx-bridge.sh up        # starts vps_node :3050 + reverse tunnel (idempotent)
bash rtx-bridge.sh test      # end-to-end: matmul through the droplet

3 · survive reboots (one admin click on Victus)

# run as Administrator once — registers a Scheduled Task that
# restarts the tunnel on boot (rtx-bridge-persist.bat):
rtx-bridge-persist.bat

// why it cannot be "down" for you

• idempotent bridge — rtx-bridge.sh up frees :3050 and kills orphans before starting, so re-running never piles up broken listeners.
• self-healing handlers — vps_node swallows client-abort (a cold gpu-mcp first call can exceed a short probe timeout; that's the client's choice, not a fault).
• point-of-truth — this page is the github.io surface; the droplet is the compute. Domain ownership (æ.store) means a live, verifiable chassis, not a parked page.