SOVEREIGN STACK

Four Pillars. Two Ports. Zero Cloud Rent.

CurXor OS is an edge appliance stack: local inference, agent engine with workspace memory and messaging gateways, ZeroMQ telemetry mesh, and Flight Command UI — outbound trades and posts isolated on eno2.

PILLAR 01

Compute

ROCm Docker inference — Ollama and vLLM on localhost. Cloud URLs rejected at startup.

64GB UMA runs massive models without sending prompts to OpenAI.

pillar-1-compute/

PILLAR 02

Engine

OpenClaw agent loop — workspace memory, skills, heartbeat scheduler, and channel gateway.

Capital, Creator, Outreach, and Forge Claws execute via dedicated bridges. Dashboard chat and Telegram, Slack, WhatsApp, or iMessage share one router synced to CCP — you configure which gateways to enable.

pillar-2-engine/

PILLAR 03

Telemetry

ZeroMQ XSUB/XPUB mesh broker for vision, motor, and swarm signals.

Low-latency local mesh — benchmark numbers pending MS-S1 MAX validation.

pillar-3-telemetry/

PILLAR 04

Flight Command

Next.js operator UI — Home hub, unified inbox, Claw workspaces, chat, and skill buttons.

Settings controls Claws, intelligence, appearance, messaging gateways, and agent runtime — wearables and health apps connect via OAuth you authorize, not a fixed vendor list.

pillar-4-dashboard/

DUAL-PORT ISOLATION

eno1 · Command Port

User LAN (10.0.0.1). Flight Command UI, captive portal, and local operator access. Firewalled from the public internet.

eno2 · Egress Port

Outbound bridge lane for agent actions — Alpaca trades, X posts, CRM webhooks, browser automation. Unplug eno2 to kill every outbound Claw instantly while local inference keeps running.

DIGITAL BRIDGES (ENO2 ONLY)

  • capital.execute_trade → Alpaca paper/live via Python bridge on eno2
  • content.publish_post → X, LinkedIn, Meta, YouTube, TikTok, and more via Python bridges on eno2
  • Channel router → Telegram, Slack, WhatsApp, iMessage — local inference, CCP inbox sync
  • Your LLM never talks to the internet — dedicated bridges and gateways handle egress

HARDWARE VALIDATION

Built for MINISFORUM MS-S1 MAX class

ROCm, UMA, and mesh benchmarks run when production silicon lands. We publish validated numbers — not marketing guesses.

  • · ROCm / UMA / mesh latency benchmarks — publish when MS-S1 MAX silicon is validated
  • · Factory USB / cloud-init install video — ships with production hardware
  • · Operator PDF bundle — export from CurXor OS docs/scripts/export-guides-pdf.sh

DESIGN PRINCIPLES

  • Sovereign edge — inference and agent control stay on localhost
  • Network isolation — eno1 (operators) and eno2 (egress) are separate concerns
  • User-configured integrations — Claws and channels you enable, not a preset vendor bundle
  • OTA golden image — flash once, iterate locally, rollback on failure

Operator documentation lives in CurXor OS under docs/guides/02-architecture.md. ← Back to pre-order