"A sovereign, air-gap-capable agentic AI platform with physical edge presence"
A homelab-scale agentic AI stack built on sovereign hardware — no cloud, no vendor lock-in, data does not leave the firewall without approval. A 120B parameter model runs on a DGX Spark as the cognitive core, a multi-agent NemoClaw crew orchestrates tasks across the infrastructure, a live RAG layer keeps the model grounded in real operational state, and a physical robot acts as a first-class edge agent in the real world.
This is not a chat wrapper. It is a goal-directed agentic system with tools, perception, and physical reach.
┌─────────────────────────────────────┐
│ OpenClaw (agent) │ ← the "brain" / agentic loop
├─────────────────────────────────────┤
│ OpenShell (policy runtime) │ ← sandboxing, guardrails, inference routing
├─────────────────────────────────────┤
│ NemoClaw (glue layer) │ ← agent onboarding, lifecycle, blueprint mgmt
├─────────────────────────────────────┤
│ vLLM + nemotron-3-super120b:a12b │ ← inference (MoE: 120B total / 12B active)
│ DGX Spark (hardware) │
└─────────────────────────────────────┘
| Component | Hostname | IP | Role |
|---|---|---|---|
| NVIDIA DGX Spark | spark-e | 10.10.12.251 | Primary inference — nemotron-super 120B via vLLM |
| NUC Gen13 × 3 | nuc-0[1-3] | 10.10.12.10[1-3] | Harvester cluster — orchestration, storage (RKE2 / Longhorn) |
| Dell XPS 9520 | wheatley | 10.10.12.252 | NemoClaw host (Ubuntu 22.04) |
| Asus i9 | jarvis | 10.10.12.250 | Local inference — RTX 4060 Ti 16GB |
| Lenovo X1 Yoga ThinkPad | blackmesa | 10.10.12.247 | Connects to external host for inference |
| Waveshare Jetbot | wall-e | 10.10.12.248 | Physical edge agent — perception + action (Jetson Nano) |
| Sophos XGS88 | — | — | Perimeter — all traffic stays inside |
| Layer | Component | Status |
|---|---|---|
| Inference (primary) | vLLM — nemotron-3-super120b (120B / 12B active MoE) | Running |
| Inference (optimized) | TensorRT-LLM — compiled engines for latency-sensitive tasks | Planned |
| Inference (fallback) | Ollama — lighter models, dev/prototyping | Available |
| Inference proxy | LiteLLM — OpenAI-compatible API on port 40000 | Running |
| Agent Runtime | NemoClaw (OpenClaw + OpenShell) | Locked |
| Agentic Framework | OpenClaw | Locked |
| RAG — Vector DB | Qdrant (K8s workload, Longhorn persistence) | Locked |
| RAG — Corpus | Live cluster state + runbooks + Jetbot sensor data | Locked |
| Orchestration | Harvester (RKE2-based HCI) | Planned |
| Identity & Auth | Authentik — OIDC/OAuth2 for all services | Locked |
| Web UI | OpenWebUI | Planned |
| Network / Perimeter | Sophos XGS88 | Existing |
Eight NemoClaw agents run the platform. Each has a defined role, a set of responsibilities, and explicit dependencies on the other crew members:
| Agent | Responsibilities |
|---|---|
| Architect | Designs solutions; writes decisions to ARCHITECTURE.md; air-gap and open-source constraints |
| Developer | IaC, application code, RAG pipelines, Jetbot integration |
| Implementer | Executes deployments — Helm, kubectl, Harvester workloads, Longhorn |
| Operations | Monitors all stack endpoints; alerts on anomalies; escalates systemic issues |
| Finance | Tracks token consumption, GPU utilization, storage growth, power; reports cost-per-work-unit |
| Security | Authentik policy auditing; air-gap egress enforcement; NeuVector; credential scanning |
| RAG Curator | Manages Qdrant corpus — ingestion, chunking, freshness, retrieval quality |
| Edge Agent | Runs wall-e's perception loop; formats sensor observations; dispatches OpenClaw actions |
Full role definitions: Agent_Roster.md
- Air-gap by default — the running system must not require internet access
- Open source, self-hostable — prefer components you can own and operate
- Agentic over chat — goal-directed loops with tools, not a chat wrapper
- Sovereignty — no cloud egress assumed; no vendor lock-in tolerated
- Decisions in the repo —
ARCHITECTURE.mdis the source of truth - Zero trust by default — every agent and service proves identity; see
Security.md
