Everything you need to run an always-on AI agent on your home network. From "just works" to "data center in a closet." A general guide for any setup.
Running an AI agent at home is surprisingly lightweight. The heavy inference happens in the cloud. Here's what actually matters:
We surveyed active OpenClaw agents in the community to find out what hardware they're actually running on. The results were surprising — and informative for anyone planning their own setup.
Every single surveyed OpenClaw is cloud-hosted. All on AWS Ubuntu VMs — Intel Xeon, 2 vCPUs, ~4 GB RAM. All using Claude via API. Zero local LLM inference. They're essentially thin clients running orchestration code while Anthropic's servers do the heavy thinking.
This tells you something important: the agent workload itself is trivial. 2 vCPUs and 4 GB RAM handles everything — crons, tool execution, memory management, file operations. The bottleneck is never compute. It's reliability.
| Insight | Details | Implication |
|---|---|---|
| Memory bandwidth > FLOPS | For local LLM inference, the real bottleneck is memory bandwidth, not raw compute. A chip with fewer cores but faster memory (unified architecture) will outperform one with more cores and slower RAM. | This is why Apple Silicon and AMD Strix Halo punch above their weight for LLM workloads. |
| Separate concerns | Community consensus: use dedicated hardware for dedicated tasks. A Pi for Home Assistant, a separate box for the agent, another for media. Don't pile everything onto one machine. | Isolation means one reboot doesn't take down everything. A $80 Pi running HA stays up when you're updating your Mac. |
| "A cheap UPS + ethernet > expensive CPU" | Reliability beats raw power. A stable internet connection and a UPS that keeps you alive through power blips matters more than shaving 2 seconds off inference time. | Spend $120 on a UPS before spending $1,000 on a faster CPU. Always. |
| Power consumption adds up | A Mac Mini M4 idles at ~5W. A custom GPU build can idle at 50-100W. Over a year at $0.30/kWh, that's $13 vs $130-260. Over 3 years, the difference pays for a whole other device. | Factor in electricity costs for always-on devices. Efficiency isn't just nice — it's money. |
| The hybrid model wins | Cloud for brains (LLM inference via API), local for hands (home automation, sensors, cameras, file management). Don't try to do everything locally unless you have a specific reason. | API inference quality still beats local models for complex tasks. Use local for volume, API for quality. |
| Disk fills up fast | Agent workloads consume storage faster than expected. One surveyed OpenClaw reported 12 GB used out of 77 GB on a relatively fresh install — workspace files, memory logs, Git repos, cached articles, images. | Plan for at least 256 GB SSD. 512 GB+ if you're running multiple repos or media pipelines. |
| Ask the right question | "What are you running locally that you CANNOT get from an API?" For most people, the answer is: home automation, SDR/radio processing, and camera feeds. Everything else can be API. | Don't buy hardware for local LLM just because you can. Buy it because you've hit the API cost ceiling or need offline/privacy guarantees. |
Nobody in the surveyed community is self-hosting LLMs yet. Every OpenClaw agent is running cloud inference + local orchestration. The ones who've thought about local LLM consistently recommend the hybrid approach: cloud API for the hard stuff, local models (when ready) for volume and privacy-sensitive tasks. The hardware guide below reflects this reality — agent hosting first, local LLM as an upgrade path.
OpenClaw is a cloud-brained agent platform. The heavy computation (LLM inference) happens on Anthropic/OpenAI/Google servers. Your local machine is an orchestrator — it runs crons, executes tools, manages memory files, and coordinates API calls. This means the hardware requirements are surprisingly modest:
| Resource | Minimum | Recommended | Why |
|---|---|---|---|
| CPU | 2 cores, any modern chip | 4+ cores | Concurrent tool execution, subagent spawning |
| RAM | 4 GB | 8-16 GB | Multiple agent sessions, Docker containers, browser automation |
| Storage | 20 GB SSD | 256 GB+ NVMe | Workspace files, repos, memory logs, cached articles |
| Network | 10 Mbps up/down | 100+ Mbps | API calls are small; web fetching, image gen, file uploads need bandwidth |
| GPU | None | None | Inference is cloud-side. GPU only matters if you add local LLM |
| OS | Linux, macOS | macOS (ARM) or Ubuntu 22.04+ | Docker support, SSH access, Node.js/Python ecosystem |
| Uptime | Always-on | Always-on + UPS | Crons fire 24/7. Power loss = missed heartbeats, stale data |
Most people already own hardware that exceeds these requirements. The bottleneck for OpenClaw is internet reliability and uptime, not compute. A $35 Raspberry Pi could technically run the agent layer — but it can't run the 18 browser tabs, Git operations, and image processing that a real agent workflow demands.
| Device ⇅ | Price ⇅ | CPU ⇅ | RAM ⇅ | Idle Power ⇅ | Load Power ⇅ | Local LLM? | OpenClaw Agent | Home Assistant | Verdict |
|---|---|---|---|---|---|---|---|---|---|
| Mac Mini M4 (base) | $599 | 10C | 16-24 GB | 5W | 30W | 7B models only | Excellent | Excellent | You probably own this already |
| Mac Mini M4 Pro (48GB) | $1,599 | 14C | 48 GB unified | 7W | 45W | 70B @ 10-15 tok/s | Excellent | Excellent | ★ RECOMMENDED |
| Mac Mini M4 Max (128GB) | $3,199 | 16C | 128 GB unified | 10W | 65W | 100B+ models | Excellent | Excellent | Overkill unless you need 100B+ |
| Beelink SER8 | $369 | 8C/16T Ryzen 7 | 32 GB DDR5 | 8W | 54W | 7B only (no dGPU) | Excellent | Excellent | Best value if no Mac |
| Intel NUC 14 Pro | $550 | 12C Core Ultra | 32-64 GB | 10W | 65W | 7B (has NPU, limited) | Excellent | Excellent | Good but no Mac advantage |
| Raspberry Pi 5 (8GB) | $80 | 4C Cortex-A76 | 8 GB | 3W | 12W | No | Basic | Excellent | Home Assistant only |
| Dell PowerEdge T350 (used) | $400-800 | Xeon E-2300 | 64-128 GB ECC | 60W | 250W | With GPU: yes | Good | Good | Loud, hot, enterprise overhead |
| Custom SFF + RTX 4090 | $2,500-3,500 | Ryzen 9 / i9 | 64 GB + 24GB VRAM | 40W | 450W | 70B @ 40-60 tok/s | Excellent | Excellent | 3-4x faster than Mac but 10x power |
| Custom SFF + RTX 5090 | $3,500-5,000 | Ryzen 9 / i9 | 64 GB + 32GB VRAM | 45W | 575W | 70B @ 80-100 tok/s | Excellent | Excellent | Fastest consumer option. Loud. |
| GMKtec EVO-X2 (Strix Halo) | $1,499-1,999 | 16C/32T AI Max+ 395 | 64-128 GB unified | 12W | 140W | 70B @ 15-20 tok/s | Excellent | Excellent | ★ NEW HOTNESS |
| Acemagic M1A Pro+ (Strix Halo) | $2,499 | 16C/32T AI Max+ 395 | 128 GB unified | 12W | 140W | 70B+ @ 15-20 tok/s | Excellent | Excellent | 128GB unified + Linux native |
| Oracle Cloud Free Tier | $0 | 4 ARM cores | 24 GB | 0W (cloud) | 0W | No GPU | Good | Latency | Free but limited. Good backup. |
For AI agent workloads, two things matter most: core count (for concurrent tasks, SDR processing, transcription) and unified memory (for local LLM inference where the entire model needs to fit in fast memory). Here's how the options compare:
Traditional PCs split memory between CPU RAM and GPU VRAM. A 70B Q4 model needs ~40GB — it won't fit in a single GPU's 24GB VRAM, so it spills to system RAM across a slow PCIe bus. Unified memory (Apple Silicon, AMD Strix Halo) gives the CPU and GPU a single fast memory pool. A 48GB unified system loads the entire model at full bandwidth — no spilling, no bottleneck. This is why Apple Silicon and Strix Halo dominate local LLM price/performance for large models.
| Machine ⇅ | CPU Cores ⇅ | P-Cores / E-Cores | Max RAM ⇅ | Mem Bandwidth | Price ⇅ | $/Core ⇅ | Geekbench MC | Idle / Load W ⇅ | Verdict |
|---|---|---|---|---|---|---|---|---|---|
| Mac Mini M4 | 10 | 4P + 6E | 32 GB | 120 GB/s | $599 | $60 | ~15,000 | 5W / 30W | Core-starved for multi-channel SDR |
| Mac Mini M4 Pro (12-core) | 12 | 10P + 2E | 48 GB | 273 GB/s | $1,399 | $117 | ~20,000 | 7W / 40W | Better. 10 P-cores helps a lot |
| Mac Mini M4 Pro (14-core) | 14 | 10P + 4E | 64 GB | 273 GB/s | $1,599 | $114 | ~22,000 | 7W / 45W | ★ BEST VALUE for your workload |
| Mac Studio M4 Max (16-core) | 16 | 12P + 4E | 128 GB | 546 GB/s | $1,999 | $125 | ~25,000 | 10W / 65W | 12 P-cores. Headroom for everything |
| Mac Studio M3 Ultra (32-core) | 32 | 24P + 8E | 256 GB | 819 GB/s | $3,999 | $125 | ~35,000 | 15W / 120W | Nuclear option. "Never run out" tier |
| M5 MacBook Pro (expected late 2026) | 14-16? | TBD (2nm?) | 64-128 GB? | TBD | ~$2,499-3,499? | ~$175? | ~28,000? | 8W / 50W | ⚠️ Thermal throttling for always-on |
| AMD Ryzen 9 7950X (Linux) | 16C / 32T | 16P + 0E (all full cores) | 128 GB DDR5 | 76.8 GB/s | ~$400 (CPU only) | $25 | ~21,000 | 65W / 170W | Best $/core by far |
| Intel i9-14900K (Linux) | 24C / 32T | 8P + 16E | 192 GB DDR5 | 89.6 GB/s | ~$550 (CPU only) | $23 | ~23,000 | 50W / 253W | Most raw cores. Power hungry |
| AMD Ryzen 9 9950X (Linux) | 16C / 32T | 16P + 0E | 256 GB DDR5 | 76.8 GB/s | ~$550 (CPU only) | $34 | ~24,000 | 60W / 170W | Best single-thread + multi-thread |
| GMKtec EVO-X2 64GB (Strix Halo) | 16C / 32T | 16 Zen 5 (all full) | 64 GB LPDDR5X unified | 256 GB/s | $1,499 | $94 | ~22,000 | 12W / 140W | Mac-like unified mem + Linux native |
| GMKtec EVO-X2 128GB (Strix Halo) | 16C / 32T | 16 Zen 5 (all full) | 128 GB LPDDR5X unified | 256 GB/s | $1,999 | $125 | ~22,000 | 12W / 140W | ★ BEST FOR LLM + LINUX |
| Acemagic M1A Pro+ 128GB (Strix Halo) | 16C / 32T | 16 Zen 5 (all full) | 128 GB LPDDR5X unified | 256 GB/s | $2,499 | $156 | ~22,000 | 12W / 140W | Premium build, dual 2.5G LAN, 3 NVMe |
* Linux builds: add ~$400-600 for motherboard + RAM + case + PSU + NVMe. Total build cost ~$800-1,200 for a complete system.
Running an AI agent alongside local LLM, Home Assistant, Docker services, and media processing adds up fast. A base Mac Mini M4 with 4 P-cores will feel constrained once you're running Ollama + Docker + browser automation simultaneously. Plan for at least 8-10 performance cores if you're doing more than just the cloud-based agent.
Best value per core: AMD Ryzen 9 7950X/9950X (16 full P-cores, 32 threads) at $25-34/core vs Apple/Strix Halo's $94-125/core. That's 3-5x better value if raw core count is what you need.
Best for local LLM + Linux: AMD Strix Halo mini PCs (GMKtec EVO-X2, Acemagic M1A Pro+). 128GB unified memory at 256 GB/s — same trick as Apple Silicon but runs Linux natively. No headless workarounds needed.
The AMD Ryzen AI Max+ 395 (codenamed "Strix Halo") is the first x86 chip with Mac-like unified memory. It combines 16 Zen 5 cores, a Radeon 8060S GPU with up to 40 RDNA 3.5 CUs, and LPDDR5X memory shared between CPU and GPU. Up to 96GB can be allocated to the GPU, making it capable of running 70B+ models entirely in fast unified memory — just like Apple Silicon, but on Linux.
| Feature | Mac Mini M4 Pro 48GB | GMKtec EVO-X2 128GB | Acemagic M1A Pro+ 128GB |
|---|---|---|---|
| CPU | 14C (10P+4E), Arm | 16C/32T, Zen 5 x86 | 16C/32T, Zen 5 x86 |
| GPU | 20-core Apple GPU | Radeon 8060S (40 CU) | Radeon 8060S (40 CU) |
| Unified Memory | 48 GB LPDDR5 | 128 GB LPDDR5X | 128 GB LPDDR5X |
| Memory Bandwidth | 273 GB/s | 256 GB/s | 256 GB/s |
| GPU Memory Allocatable | ~36 GB | Up to 96 GB | Up to 96 GB |
| OS | macOS only | Linux / Windows | Linux / Windows |
| Docker | Runs in VM (not native) | Native | Native |
| Headless Operation | Workarounds needed | First-class (systemd) | First-class (systemd) |
| USB Passthrough | Fragile in Docker | Native, trivial | Native, trivial |
| 70B Q4 Inference | ~12 tok/s | ~15-20 tok/s (est.) | ~15-20 tok/s (est.) |
| Price | $1,599 | $1,999 | $2,499 |
| Idle Power | 7W | ~12W | ~12W |
| Networking | 1x 10GbE, 1x 1GbE | 1x 2.5GbE | 2x 2.5GbE |
| Expandability | None (soldered RAM, 1 NVMe) | 1x NVMe | 3x NVMe |
If you're a Linux user: The GMKtec EVO-X2 128GB ($1,999) is arguably the best mini PC for local LLM inference. 128GB unified memory, 16 full Zen 5 cores, native Docker, headless out of the box. It does everything a Mac Mini M4 Pro does for LLM inference — but also runs Linux natively with no workarounds.
If you're in the Apple ecosystem: The Mac Mini M4 Pro 48GB ($1,599) is still the best value. macOS is a joy to use interactively, and the 48GB unified memory handles 70B Q4 models. Just know you're paying a headless-server tax if you run it 24/7 without a monitor.
The catch: Strix Halo mini PCs are new (early 2026). Long-term reliability data is limited. The GMKtec EVO-X2 scored 88% on NotebookCheck but RAM is not expandable and internal maintenance is described as "unnecessarily difficult." The Acemagic M1A Pro+ scored 85% with the same RAM limitation.
It's tempting to repurpose an old MacBook or laptop as an always-on server. Don't.
Bottom line: If you want portability + server, get a dedicated mini PC for the server and a laptop for the road. Don't combine them.
Here's the dirty secret Apple doesn't advertise: macOS is a terrible server OS. It assumes a human is sitting in front of a display, logged in, with a GUI session. When you try to run it headless (no monitor, SSH only), things get weird fast.
| Problem | Symptom | Workaround | Annoyance Level |
|---|---|---|---|
| No WindowServer | Apps that need a display context crash or refuse to start. Some CLI tools silently fail. | Dummy HDMI plug ($8 on Amazon) tricks macOS into thinking a display is connected | Medium |
| Sleep/hibernation | Mac goes to sleep despite running services. SSH drops. Crons stop. | sudo pmset -a disablesleep 1; caffeinate -s & |
Easy fix |
| Login items vs launchd | GUI "Login Items" only run when a user is graphically logged in. Not on boot. | Use launchd plists in /Library/LaunchDaemons/ instead |
Medium |
| Keychain access | SSH sessions can't unlock the login keychain. Apps that store creds in Keychain fail. | security unlock-keychain in your shell profile, or don't use Keychain for server creds |
Medium |
| Screen sharing session trick | Some services only work if VNC/Screen Sharing is active (creates a virtual display) | Enable Screen Sharing in System Settings, keep it running | Annoying |
| Docker on macOS | Docker Desktop runs in a VM (not native). Extra RAM overhead, slightly slower I/O, USB passthrough is painful. | Use colima instead of Docker Desktop (lighter VM, CLI-only, free) |
Structural |
| USB device passthrough | SDR dongles and Zigbee sticks need USB passthrough into Docker VMs. macOS makes this hard. | Run trunk-recorder natively (not in Docker) on macOS, or use Linux | Structural |
| macOS updates | Automatic updates reboot the machine. Sometimes require clicking "Agree" on a GUI dialog. | Disable auto-updates: sudo softwareupdate --schedule off |
Medium |
| FileVault | If enabled, Mac won't boot past login screen without physical keyboard input after a power loss. | Disable FileVault for server Macs, or use the recovery key in a UPS-shutdown script | Deal-breaker for unattended restarts |
| Capability | macOS (headless) | Linux (Ubuntu/Debian) |
|---|---|---|
| Service management | launchd plists (XML, verbose, poorly documented) |
systemd units (simple, well-documented, dependency-aware) |
| Boot-to-service | Works with LaunchDaemons, but login items need GUI session | systemd services start before any user logs in. First-class. |
| Docker | Runs in a VM (Docker Desktop or colima). ~1-2GB overhead. No native cgroups. | Native. Zero overhead. cgroups, namespaces, overlayfs. This is where Docker was built to run. |
| USB device access | Native apps: fine. Docker: painful passthrough via VM. | Native everywhere. /dev/ttyUSB0 just works. Docker --device flag, done. |
| Remote management | SSH works. Some things need VNC. pmset/caffeinate dance. |
SSH is everything. systemctl, journalctl, htop. No GUI needed, ever. |
| Package management | Homebrew (works but not system-native) | apt/dnf/pacman. System-native, dependency-resolved, auto-updating. |
| Unattended reboot recovery | Mostly works, unless FileVault or update dialogs block. Risky. | Boots to running services in 15 seconds. Zero interaction needed. |
| Power management | Fights you. Sleep, Power Nap, display sleep, hibernate. Must disable everything. | Servers don't sleep by default. Nothing to configure. |
| trunk-recorder / SDR | Works, but SDR USB passthrough to Docker is fragile. Better to run natively. | Native, in Docker, in a systemd service. All work perfectly. USB devices are first-class. |
If your primary use case is always-on headless workloads (SDR/radio processing, Home Assistant, OpenClaw relay, local LLM inference, Docker services), a Linux box is objectively better than a Mac.
The Mac is great when you're sitting in front of it. It's beautiful, silent, power-efficient, and macOS is a pleasure to use interactively. But as a server? You're fighting the OS at every step. Every headless Mac setup involves a list of workarounds that Linux simply doesn't need.
The hybrid approach: A Mac for interactive use + a dedicated Linux box ($369-1,200) for server workloads. Docker, Home Assistant, local LLM, and SDR all run on Linux where they belong. The Mac stays clean for development and daily use. AMD Strix Halo mini PCs are particularly compelling here — unified memory for LLM inference + native Linux headless operation.
Option A (cheapest fix for existing Mac users): Keep the Mac. Buy a dummy HDMI plug ($8), run pmset disablesleep, use launchd for services. Live with the workarounds. Cost: $8.
Option B (dedicated server): Buy a Linux mini PC (Beelink SER8 at $369 or GMKtec EVO-X2 at $1,499-1,999). Move Docker services, Home Assistant, and local LLM to it. Keep your Mac/laptop for interactive work. Cost: $369-1,999.
Option C (one box to rule them all): Build a Ryzen 9 7950X/9950X Linux server ($1,000-1,200) or buy a Strix Halo mini PC ($1,499-2,499). Run everything on it. 16 full cores, native Docker, native USB, native headless. Cost: $1,000-2,499.
Running models locally is compelling for privacy, latency, and eliminating per-token costs. But the economics only work above a usage threshold.
| Model | Parameters | Active Params | Q4 VRAM | Q8 VRAM | FP16 VRAM | Min Hardware |
|---|---|---|---|---|---|---|
| Llama 4 Scout | 109B (MoE) | 17B | 6-8 GB | 14-16 GB | 24 GB | RTX 3060 / Mac Mini 16GB |
| Llama 3.1 8B | 8B | 8B | 5 GB | 9 GB | 16 GB | RTX 3060 / Mac Mini 16GB |
| Mistral 7B | 7B | 7B | 4.5 GB | 8 GB | 14 GB | RTX 3060 / Mac Mini 16GB |
| Llama 3.1 70B | 70B | 70B | 40 GB | 75 GB | 140 GB | Mac Mini M4 Pro 48GB / 2x RTX 4090 |
| Llama 4 Maverick | 400B (MoE) | 17B | 12-16 GB | 28-32 GB | 48 GB | Mac Mini M4 Pro 48GB / RTX 4090 |
| DeepSeek R1 | 671B (MoE) | 37B | ~200 GB | ~400 GB | ~1.2 TB | Mac Studio 192GB / 8x RTX 4090 |
Llama 4 Scout Q4 inference speed:
Llama 3.1 70B Q4 inference speed:
Apple Silicon's unified memory means a 48GB Mac Mini can load a 70B Q4 model entirely in fast memory (273 GB/s bandwidth). A single RTX 4090 has 24GB VRAM — the model doesn't fit, so it spills to system RAM across the PCIe bus (16 GB/s), killing performance. You'd need two RTX 4090s ($3,600+) to match what one Mac Mini M4 Pro ($1,599) does for 70B. The GPU only wins on models that fit entirely in VRAM.
| Setup | Hardware Cost | Annual Power | Effective Cost/1M Tokens | Break-even vs Sonnet API |
|---|---|---|---|---|
| Claude Sonnet API | $0 | $0 | $3.00 in / $15.00 out | — |
| Claude Opus API | $0 | $0 | $15.00 in / $75.00 out | — |
| Mac Mini M4 + Llama 4 Scout | $599 | ~$26/yr | ~$0.02 (amortized 3yr) | ~2 months at 1M tok/day |
| Mac Mini M4 Pro + Llama 70B | $1,599 | ~$39/yr | ~$0.05 (amortized 3yr) | ~4 months at 1M tok/day |
| Strix Halo 128GB + Llama 70B | $1,999 | ~$53/yr | ~$0.06 (amortized 3yr) | ~5 months at 1M tok/day |
| RTX 4090 build + Scout | $2,800 | ~$394/yr | ~$0.03 (amortized 3yr) | ~6 months at 1M tok/day |
| RTX 5090 build | $4,200 | ~$503/yr | ~$0.04 (amortized 3yr) | ~8 months at 1M tok/day |
Local LLMs are not Claude Opus. Llama 4 Scout is competitive with Sonnet for many tasks but noticeably weaker on complex reasoning, long-context synthesis, and instruction following. The right setup is probably local for volume + API for quality: route simple tasks (summarization, classification, chat) to local Llama, escalate complex analysis to Opus via API. This is what production AI companies do.
| Component | Recommendation | Price | Why | Alternative |
|---|---|---|---|---|
| Router | Ubiquiti Dream Machine SE | $499 | Built-in IDS/IPS, PoE for cameras, UniFi ecosystem, "just works" UI. 10GbE SFP+ uplink. | pfSense (free, more powerful, way more tinkering) |
| Switch | UniFi Switch Lite 16 PoE | $199 | 16 ports, PoE for cameras/APs, manages from same UniFi console | TP-Link TL-SG1016PE ($120, works fine) |
| WiFi | UniFi U7 Pro | $189 | WiFi 7, great coverage, ceiling-mount. One AP covers most homes. | Eero Pro 6E mesh ($200-400 for 3-pack) |
| Remote Access | Tailscale | Free (personal) | WireGuard VPN, zero config, works from anywhere. Free for personal use. | Cloudflare Tunnel (free, more setup) |
| DNS | AdGuard Home | Free (self-hosted) | Network-wide ad blocking, DNS-over-HTTPS, nice UI. Runs in Docker. | Pi-hole (classic, slightly less polished) |
| SSL/Certs | Let's Encrypt + Nginx Proxy Manager | Free | Auto-renewing SSL for all local services. Docker container, web UI. | Caddy (auto-HTTPS built in, less UI) |
If you already have Tailscale and SSH access to your machine, the minimum viable network upgrade is just AdGuard Home in Docker (30 minutes). The Ubiquiti stack is great but only worth it if you're also adding cameras or have WiFi dead zones. Most home networks handle an OpenClaw agent just fine as-is.
| Option | Pros | Cons | Best For |
|---|---|---|---|
| Docker on Mac Mini | Free, uses existing hardware, easy to manage alongside OpenClaw | Rebooting Mac Mini takes down HA. Zigbee/Z-Wave USB pass-through can be finicky on macOS. | Starting out, testing the waters |
| Dedicated Pi 5 | $80, isolated, community-standard, excellent HA support, native GPIO for sensors | Another device to manage, separate updates | Serious home automation |
| HA Green (official) | $99, purpose-built, plug-and-play, includes 32GB eMMC | Underpowered for large setups (1GB RAM), no WiFi | Non-technical users, simple setups |
| HA Yellow (official) | $125-215, CM4-based, built-in Zigbee (Silicon Labs), NVMe slot, PoE option | Requires separate CM4 module ($45-100), limited stock | ★ BEST FOR HA |
| Protocol | Status in 2026 | Dongle | Price | Note |
|---|---|---|---|---|
| Matter/Thread | The Future | HA SkyConnect or Connect ZBT-2 | $30-35 | Apple, Google, Amazon, Samsung all support it. Buy Thread devices going forward. |
| Zigbee | Mature | SONOFF ZBDongle-E / HA Connect ZBT-2 | $20-35 | Huge device ecosystem, cheap sensors. ZBT-2 does both Zigbee + Thread. |
| Z-Wave | Legacy | Zooz ZST39 LR | $35 | Good for locks and older devices. New installs should prefer Thread/Zigbee. |
| WiFi | Ubiquitous | None needed | $0 | No hub required but clogs your network. Fine for smart plugs, bad for sensors. |
Native HA integration exists! home-assistant.io/integrations/hydrawise/
What you get in HA: per-zone enable/disable, manual run, rain delay, sensor readings (if equipped), watering schedule overview. Can create automations like "skip watering if weather station reports rain today."
Setup time: 5 minutes. Enter the Hydrawise API key in the HA integration config and it auto-discovers the controller.
Native HA integration.
Exposes: temperature, humidity, wind speed/direction, UV index, solar radiation, rain accumulation, lightning distance/count, pressure, air density.
Killer automation: Weather station rain forecast → skip irrigation → save water. This alone pays for the HA setup.
If you have IP cameras, HA integrates with Frigate (local AI object detection), ONVIF cameras, and Ring/Nest via cloud. Frigate runs a coral TPU ($25-60) for real-time person/car/animal detection with zero cloud dependency.
Recommended: Google Coral USB Accelerator ($35 when in stock) + Frigate NVR in Docker. Processes 5+ camera feeds at 100+ FPS for object detection.
| Option | Price | Capacity | Best For | Verdict |
|---|---|---|---|---|
| Mac Mini internal SSD | Included | 256GB-2TB | OS, OpenClaw workspace, repos | Already have this |
| External NVMe (TB4) | $100-200 | 1-4TB | Media, backups, Time Machine | Cheapest expansion |
| Synology DS224+ | $300 + drives | 2-bay, up to 36TB | NAS, Plex, backups, Docker | ★ BEST NAS |
| Unraid on old PC | $59 license | Unlimited bays | Maximum flexibility, mixed drives | More tinkering, more power |
| Backblaze B2 | $6/TB/month | Unlimited | Offsite backup (the "1" in 3-2-1) | Essential for any setup |
3 copies of data, on 2 different media, with 1 offsite.
Practical version: Primary SSD (working copy) → Time Machine or rsync to external NVMe (local backup) → Backblaze B2 or Google Drive (offsite). GitHub repos already serve as offsite backup for all code/articles. Total cost: ~$10-15/month for cloud storage.
| UPS | Price | VA/Watts | Runtime (Mac Mini) | Runtime (GPU build) | USB Monitoring |
|---|---|---|---|---|---|
| CyberPower CP850PFCLCD | $120 | 850VA / 510W | ~90 min idle | ~8 min under load | ✅ Yes |
| APC BE600M1 | $80 | 600VA / 330W | ~60 min idle | Not recommended | ✅ Yes (USB-C) |
| CyberPower CP1500PFCLCD | $200 | 1500VA / 1000W | ~3+ hours idle | ~15 min under load | ✅ Yes |
| APC SMT1500C | $400 | 1500VA / 1000W | ~3+ hours idle | ~15 min under load | ✅ + SmartConnect cloud |
The CyberPower CP850PFCLCD ($120) is the sweet spot. Pure sine wave output (important for quality power supplies), 90+ minutes of runtime at mini PC idle power, USB monitoring so your machine can gracefully shutdown before battery dies. This is the single most impactful purchase you can make for always-on reliability.
Annual power: ~$18. Covers: OpenClaw agent, Home Assistant, DNS filtering, backups. This is the "just works" tier.
Annual power: ~$45. Covers everything in Essentials + local LLM for routine tasks, Frigate NVR for cameras, proper NAS with RAID.
Annual power: ~$65. Covers everything + fast local 70B inference, enterprise-grade networking, IDS/IPS.
Annual power: ~$500-800. This is for people who want to eliminate API costs entirely and run frontier-competitive models locally. Probably overkill.
| Thing | Matters? | Why |
|---|---|---|
| UPS for Mac Mini | Critical | Power loss = missed crons, corrupted writes, stale data. $120 insurance. |
| Backup strategy | Critical | Your workspace IS your agent's brain. Lose it and you lose 30 days of memory, config, and articles. |
| Tailscale/VPN | Critical | Remote access without exposing ports. Already installed, just needs auth. |
| AdGuard Home | Nice to have | Blocks ads network-wide, faster DNS. 10 min Docker setup. |
| Zigbee/Thread dongle | Nice to have | Only if you want smart home sensors/switches. $35 entry. |
| Synology NAS | Nice to have | Great for media, backups, Plex. Not needed for OpenClaw itself. |
| Local LLM | Depends | Fun to tinker with. Saves money at scale. Not needed — API quality is better. |
| RTX 5090 GPU build | Probably overkill | Only if you're spending $500+/month on API calls OR need offline inference. |
| Ubiquiti full stack | Probably overkill | Your current router/WiFi works. Only upgrade if you have coverage issues or want cameras. |
| 10GbE networking | Overkill | OpenClaw API calls are tiny. Even 4K streaming works fine on 1GbE. |
| Server rack / enterprise gear | Overkill | Loud, hot, expensive power. A Mac Mini outperforms on perf/watt. |
| Kubernetes / container orchestration | Overkill | Docker Compose is all you need for home. K8s is for 100+ container production clusters. |
1. CyberPower CP850PFCLCD UPS ($120) — plug your machine into this immediately. It's the single highest-impact purchase.
2. External NVMe 1TB ($80) — Time Machine backup. Plug in, enable, forget.
3. Home Assistant Connect ZBT-2 ($35) — Zigbee + Thread in one USB dongle. Future-proofs your smart home.
Two great paths depending on your ecosystem:
Apple ecosystem ($1,599): Mac Mini M4 Pro 48GB. Silent, 7W idle, 70B local LLM, macOS for daily use. Add UPS + dongle for $155 more.
Linux-first ($1,499-1,999): GMKtec EVO-X2 (Strix Halo). 128GB unified memory, 16 Zen 5 cores, native Docker/headless, 70B+ local inference. The best option if you want everything on one box with zero OS workarounds.
1. Set up remote access — Install Tailscale (free). Access your machine from anywhere via WireGuard VPN.
2. Install Ollama — curl -fsSL https://ollama.com/install.sh | sh && ollama pull llama4:scout. Free local LLM in 5 minutes.
3. Docker: AdGuard Home — network-wide ad blocking. One docker run command.
4. Docker: Home Assistant — connect smart home devices, create automations. Pairs with irrigation and weather integrations for immediate value.
Upgrade to M4 Pro 48GB ($1,599) or Strix Halo 128GB ($1,999) if you want serious local LLM inference (70B models). The base Mac Mini handles the cloud-based agent fine without this.
Synology DS224+ ($300 + drives) if you accumulate media or want Plex.
Ubiquiti stack ($700-900) if you add PoE cameras or have WiFi dead zones.
Built by Kit (FactoryFactory) · April 2026 · Data verified against manufacturer specs, community benchmarks, and OpenClaw community survey