Hardware matrix

Read this first#

You can run an OpenClaw Gateway on a wide spectrum of hardware — from a Raspberry Pi sitting on your desk to an Azure VM in a datacenter. The trade-offs are real but knowable. This page is the cheat-sheet I wished I had before picking where to start.

TL;DR: if you have a recent-ish Mac (M1 or newer), use it. If you don’t, a Raspberry Pi 5 with 8GB is the strongest “buy it and forget it” pick. Everything else is a variation on those two themes.

The matrix#

Hardware	RAM	OpenClaw fit	Voice features	Cost (USD-ish)	Best for
Mac M1 / M2 / M3 / M4	8GB+	Excellent — runtime + local model possible	Yes (Voice Wake + Talk Mode native)	$1,000+ (you probably already own one)	Daily driver. Best feature parity, easiest install.
Linux desktop (mid-range)	16GB	Excellent — runtime + local model	Limited (text channels only — no native voice node)	$400–$1,500	If you live on Linux. WSL2 on Windows works the same way.
Intel NUC / mini PC	16GB	Excellent — silent always-on home server	Limited	$400–$700	”Set and forget” on your desk, fanless or quiet.
Raspberry Pi 5 (8GB)	8GB	Good — runtime fine; small local model only	None (no voice node)	$80 + power supply + SD card ≈ $130	Cheap always-on. Most people’s first OpenClaw box.
Raspberry Pi 4 (4GB or 8GB)	4GB / 8GB	OK with caveats — see Pi page	None	$40–$70	Already-have-one starter. 4GB is tight.
Old laptop (recycle)	4GB+	Depends on age — anything from ~2018+ likely fine	Maybe (Mac yes, others limited)	Free if you have one	Best value if you’ve got one collecting dust.
Azure VM (B-series small)	2GB+ (B2s = 4GB)	Workable for text channels; no local model	None	~$40–$100/month	Production-style deployment. See §2.5 Azure.
Azure Container Apps	configurable	Workable for stateless restarts; persistence needs care	None	scales with use	Serverless-ish path. See §2.5 Azure.
Old gaming desktop with GPU	16GB+ + GPU	Excellent if you want to run local models	Limited	already have it	If you’ve already invested in local LLMs (Ollama et al), the Gateway can drive them locally.
NAS (Synology / TrueNAS / Unraid)	depends on model	Possible via Docker — see §2.7 Docker	None	already have it	Quiet 24/7 home server you already pay for in electricity.

What “fit” actually means#

The Gateway runtime is Node 24 (or 22.14+). It needs:

• ~200–500MB of RAM at idle (estimated, one Gateway, one agent — not measured by us yet)
• More if you load skills / MCP servers — each loaded skill or MCP adds context
• Network connectivity to your model provider — unless you’re running local models, every message round-trip hits Anthropic / OpenAI / etc.
• Persistent disk for sessions JSONL (~/.openclaw/agents/<agentId>/sessions/) and workspace files
• The ability to keep a process alive — launchd on macOS, systemd on Linux, your platform’s equivalent on Windows

So the runtime itself is light. What grows your hardware needs is what you connect to it.

The table below is illustrative — sourced inferences, not measurements. Local-model RAM/disk numbers come from the model providers’ published specs (Ollama / Hugging Face). The Gateway-side figures are educated guesses pending a real measurement pass; treat them as ranges, not guarantees.

Workload	Extra RAM needed	Extra disk needed
Just the Gateway, text channels, API-based models	Negligible	Negligible
Adding 5+ MCP servers (filesystem, GitHub, browser)	+100–500MB (estimated)	Negligible
Long sessions accumulating JSONL transcripts	Negligible	~10MB / month / active session (estimated)
Local model via Ollama (e.g. Llama 3.1 8B)	+6GB at minimum (model spec)	+5–10GB per model (model spec)
Local model larger (70B+)	+50GB+ and you need a decent GPU (model spec)	+40GB+ per model (model spec)
Voice features (TTS, ASR)	+500MB–1GB (estimated)	+1GB (estimated)

So if you want API-only models + text channels, almost any modern computer works. If you want local models, plan for the model’s footprint not the runtime’s.

Why this matters#

A common mistake is to over-spec the box because “AI = expensive hardware.” That’s true if you’re training models. It’s not true if you’re running an agent runtime that calls a hosted API. A $130 Pi 5 can do the same agent work as a $3,000 Mac, as long as the model is hosted.

The decision really comes down to:

1. What’s always-on? (Pi / NUC / NAS / VM are 24/7. A laptop closes.)
2. Do you want voice? (Mac/iOS/Android only — see §1.4 drawback #6.)
3. Do you want a local model? (Then RAM and possibly GPU dominate; otherwise hosted API is fine.)
4. What’s the threat surface? (Always-on internet-facing changes the security calculus — see §6.1 Self-hosting checklist.)

Recommended starter box#

If you don’t have anything yet and want the cheapest credible path to “I have an OpenClaw and it works”:

• Raspberry Pi 5, 8GB ($80)
• 64GB or 128GB MicroSD card or NVMe (faster) ($15–$30)
• Official 27W USB-C power supply ($12)
• A case with a fan (the Pi 5 will throttle without one) ($15)
• An ethernet cable to your router (faster + more reliable than wifi)

Total: ~$130. Boot Raspberry Pi OS Lite (or Ubuntu Server). Follow §2.6 Raspberry Pi. Connect the channels you actually use. Done.

What we are NOT going to claim#

We have not benchmarked OpenClaw on every box in the table. RAM/CPU floors are inferred from Node runtime requirements + architectural docs + community discussion. Specific timing numbers (cold-start latency, memory-at-rest) need actual runs to confirm. Sush will run on his M2 MacBook (Mac column promotes to tried) and a Pi 5 (Pi column promotes to tried) — those two will get real numbers first.

What to read next#

• You picked your hardware. → §2.2 Decision tree (planned for P0b) or jump to the right setup page below
• §2.3 Laptop quick-start — Mac / Linux desktop / Windows WSL2
• §2.6 Raspberry Pi — Pi 4 / Pi 5
• §2.5 Azure — Container Apps + VM paths
• §2.7 Docker — packaging method that wraps several of the above
• §6.1 Self-hosting checklist — turn the box into a credible deployment