A multi-node homelab built around local AI inference, self-hosted services, smart home automation, and an AI assistant control plane — fully declared as code.

The driving philosophy: everything runs locally. LLMs, image generation, DNS, media automation, home automation — no cloud dependencies, no subscriptions, no data leaving the network.

graph TB
    subgraph network["🌐 Home Network — 192.168.68.0/24"]
        direction TB

        subgraph strix["🤖 GMKTEC EvoX2 — Strix Halo · 192.168.68.78"]
            direction LR
            ollama["Ollama\nLocal LLM Inference"]
            webui["Open WebUI\nChat Interface"]
            comfyui["ComfyUI\nImage Generation"]
        end

        subgraph ubuntupn["🦞 Dell Laptop — Ubuntu · 192.168.68.82"]
            direction LR
            openclaw["OpenClaw Gateway\nAI Assistant Control Plane"]
            telegram["Telegram Bot\nChat Channel"]
        end

        subgraph proxmox["🖥️ HP EliteDesk 800 G4 — Proxmox VE · 192.168.68.50"]
            direction LR
            pihole["Pi-hole\nDNS · .51"]
            media["Media Stack\nSonarr · Radarr · qBit · .53"]
            ha["Home Assistant\nZigbee · Smart Home · .52"]
        end

        subgraph nas["💾 TerraMaster F4-424 Max — Unraid · 192.168.68.72"]
            direction LR
            nfs["NFS Shares"]
            storage["16 TB Raw Storage\n2x 8TB WD HDDs"]
        end
    end

    webui --> ollama
    openclaw -- "SSH · AI inference" --> strix
    openclaw -- "SSH · VM/CT mgmt" --> proxmox
    openclaw -- "SMB · File storage" --> nas
    strix -- "NFS Mount" --> nfs
    proxmox -- "NFS Mount" --> nfs
    media -- "Downloads & Media" --> storage

    style strix fill:#1a1a2e,stroke:#e94560,color:#fff
    style ubuntupn fill:#1a1a2e,stroke:#f5a623,color:#fff
    style proxmox fill:#1a1a2e,stroke:#0f3460,color:#fff
    style nas fill:#1a1a2e,stroke:#16213e,color:#fff
    style network fill:#0d1117,stroke:#30363d,color:#fff

Cloud AI APIs are convenient. But running models locally means:

• Privacy — conversations, images, and prompts never leave my network
• No rate limits — inference runs as fast as the hardware allows
• No recurring costs — after hardware investment, inference is free
• Real learning — configuring GPU drivers, memory allocation, and model optimization teaches you things that API calls never will

The Strix Halo's 128 GB of unified memory is the centerpiece — it can load large language models (70B+ parameter) and run Stable Diffusion image generation with ~96 GB VRAM allocation, producing images in roughly 7 seconds.

OpenClaw is an open-source personal AI assistant that runs on your own devices. It acts as the control plane for my homelab, connecting all machines through a unified interface accessible via Telegram.

The gateway runs on the Dell laptop (ubuntupn) and has SSH access to the Strix Halo and Proxmox, plus SMB access to the NAS — giving it the ability to manage infrastructure, run AI workloads, and access stored files, all from a chat interface.

Rather than giving OpenClaw root access everywhere, I created a dedicated low-privilege adrian user across all machines with carefully scoped permissions:

Ollama's network port on the Strix Halo is firewalled to only accept connections from the OpenClaw gateway:

sudo ufw allow from 192.168.68.82 to any port 11434

The OpenClaw dashboard requires HTTPS or localhost for WebSocket connections, so LAN access uses an SSH tunnel:

ssh -L 18789:127.0.0.1:18789 adrian@192.168.68.82

Telegram DMs use pairing-based security — unknown senders receive a code that must be manually approved before the bot responds.

See openclaw/openclaw.json.example for the obfuscated configuration template.

The Strix Halo uses AMD's gfx1151 GPU architecture, which had no official ROCm support at the time of setup. Getting local AI inference working required:

• Community-built PyTorch wheels — official AMD wheels didn't support gfx1151, so I used builds from community developers targeting this specific architecture
• Custom kernel parameters — tuned amdgpu.gttsize and ttm.pages_limit to allocate ~96 GB of the 128 GB unified memory to GPU workloads, leaving ~32 GB for the OS and OpenClaw operations
• ROCm environment overrides — set HSA_OVERRIDE_GFX_VERSION and PYTORCH_HIP_ALLOC_CONF to make the stack recognize and properly utilize the hardware

This is the kind of problem you only encounter when you're running real workloads on hardware that's ahead of the software ecosystem — exactly where production edge-AI deployments often land.

Each machine has a clear role: NAS handles storage, Proxmox hosts general services in isolated containers, the Strix Halo is dedicated to GPU-heavy AI workloads, and the Dell laptop runs the always-on AI assistant gateway. This mirrors how production infrastructure separates compute, storage, and orchestration — and makes troubleshooting significantly easier.

The OpenClaw integration required creating a consistent security model across four machines running three different operating systems (Ubuntu, Proxmox, Unraid). Each machine has a dedicated service user with only the permissions needed — SSH key auth where possible, custom Proxmox API roles with granular privilege sets, and SMB shares scoped to a single directory. No machine trusts the assistant with more than it needs.

The media automation stack routes all traffic through a VPN at the container level, so only download traffic is tunneled while the rest of the network operates normally. This was configured using split-tunneling within the LXC container.

Home Assistant runs as a dedicated VM on Proxmox with Zigbee device integration planned. The goal is full local control of lighting, sensors, and automation — no cloud hubs, no vendor lock-in.

(Zigbee coordinator and device integration in progress)

homelab/
├── README.md                          # You are here
├── strix-halo/
│   ├── docker-compose.yml             # Open WebUI
│   ├── comfyui.service                # systemd unit for ComfyUI
│   ├── setup.sh                       # Reproducible setup script
│   ├── .env.example                   # Environment variables template
│   └── README.md                      # Strix Halo specific docs
├── openclaw/
│   ├── openclaw.json.example          # Obfuscated config template
│   ├── openclaw.service               # systemd user service
│   └── README.md                      # OpenClaw setup & security model
├── proxmox/
│   ├── pihole/
│   │   └── docker-compose.yml
│   ├── media-stack/
│   │   ├── docker-compose.yml         # Sonarr, Radarr, Prowlarr, qBit
│   │   └── .env.example
│   ├── home-assistant/
│   │   └── README.md                  # VM setup notes
│   └── README.md                      # Proxmox host config & LXC setup
├── nas/
│   └── README.md                      # Unraid config, shares, NFS exports
└── docs/
    ├── hardware.md                    # Detailed hardware specs
    ├── network.md                     # IP assignments, DNS, VPN
    └── recovery.md                    # Disaster recovery runbook

• Zigbee coordinator setup with Home Assistant
• k3s cluster across homelab nodes
• Voice processing pipeline (Whisper + Piper for STT/TTS)
• Vaultwarden for self-hosted password management
• Frigate NVR for camera monitoring
• Comprehensive backup automation to NAS
• GitOps workflow with ArgoCD

MIT