Self-Hosted AI Agent

Do you need a Mac mini for OpenClaw?

No. OpenClaw runs on macOS, Linux or Windows and needs only about 1–2 GB of RAM, because the AI model normally runs in the cloud, so a computer you already own is usually enough. A Mac mini is worth buying if you want iMessage or other Mac-only features, or want to run AI models on the machine itself. For iMessage, the base M6 model with 16 GB ($899 as of September 2026) is plenty. For local models, memory is what counts: get 24 GB or more.

By Tyler Mayberry, creator of ibara and founder of Cirlet · Updated 29 September 2026

The short verdict

Buy a Mac mini only if one of the reasons below applies to you. Otherwise, start with hardware you have. OpenClaw's own FAQ answers "Do I have to buy a Mac Mini to install this?" with "No", and says a small server, a home server or a Raspberry Pi-class box works too.

A Mac mini is worth it if:

What you already have is enough if:

Not sure which applies? The free tool on the home page, "Which computer should run your agent?", walks through it.

What OpenClaw actually needs

OpenClaw's main program (the "Gateway") is a Node.js app, so it needs very little. As of September 2026, the official docs list:

Two limits catch people with old hardware. Current Node versions need macOS 13.5 or later, so a Mac stuck on macOS 11 to 13.4 can no longer run the OpenClaw Gateway. And there are no supported Node builds for 32-bit ARM Linux, so a Raspberry Pi needs a 64-bit operating system.

The heavy lifting is the AI model. With a cloud model, your computer just passes messages back and forth and runs tools. That's why a 1 GB server can work. Running the model locally is a different job, covered below.

Why so many people bought Mac minis

People bought Mac minis because they're small desktop Macs that can stay on all the time, with memory options large enough for local models. In February 2026, Business Insider reported that OpenClaw had people buying Mac minis faster than Apple could keep up, with 24 GB and 32 GB configurations showing waits of several weeks.

The appeal is real. Apple lists the current M6 Mac mini at 4 W idle, and it runs macOS, which iMessage needs. But if you'll use a cloud model and a chat app other than iMessage, OpenClaw only needs a computer that stays on, and almost any modern one will do.

What only a Mac can do

iMessage is the main feature that needs a Mac. OpenClaw connects to iMessage through a tool called imsg, which must run on a Mac signed in to Messages, with Full Disk Access and Automation permissions granted.

The OpenClaw Mac app adds a menu bar icon, notifications, voice input and tools for controlling the Mac. You can get those by pairing a Mac you already own as a "node" to a Gateway running somewhere else.

Which Mac mini to buy

If you don't need local models, buy the cheapest one. OpenClaw itself fits easily in 16 GB, and 256 GB of storage is far more than it uses. Apple's US prices as of September 2026:

Mac miniPriceGood for
M6, 16 GB, 256 GB$899OpenClaw with cloud models, iMessage, Mac apps. Small local models only.
M6, 16 GB, 512 GB$1,099Same as above, with room for large downloads.
M6, 24 GB, 512 GB$1,299Mid-size local models (around 12 billion parameters).
M6, 32 GB, 1 TB$1,799The largest models in OpenClaw's managed local setup (around 27–30 billion parameters).
M5 Pro, 64 GB, 1 TB$3,199Bigger local models, run through your own model server.

Those memory tiers come from OpenClaw's built-in local model setup, which picks a model by how much memory you have: 8 GiB for its smallest recipe, 16 GiB for a 9-billion-parameter model, 24 GiB with graphics acceleration for a 12-billion-parameter model, and 32 GiB for its largest. On Apple silicon it uses the graphics chip through Metal, which is why Macs do well here. OpenClaw says these are minimums, not promises of speed.

Local models have a catch. OpenClaw's FAQ answers "Is a local model OK for casual chats?" with "Usually no", because OpenClaw needs long context and strong safety, and small or heavily compressed models raise the risk of prompt injection. The docs show how to keep a cloud model as the main one with a local model as a fallback. See what OpenClaw costs to run to compare against cloud model bills.

A used Mac mini can be a good deal if it runs macOS 13.5 or later. Prefer Apple silicon: Apple lists the 2020 M1 Mac mini at 6.8 W idle, against 19.9 W for the 2018 Intel model.

The alternatives compared

Any of these runs OpenClaw well with a cloud model; they differ in cost, power, Mac features and local models.

OptionUpfront costIdle poweriMessageLocal modelsWatch out for
Computer you already use$0VariesOnly if it's a MacDepends on memoryPauses when it sleeps, restarts or updates
Old laptop (spare)$0VariesOnly if it's a Mac on macOS 13.5+Depends on memoryLid-close sleep; old Macs may be too old
Office-surplus desktop or mini PCUsed market priceAbout 9–12 W (1-litre office PCs)NoSmall models on the processor onlyCheck it's 64-bit with 2 GB+ RAM
Raspberry Pi 5$35–65 board (1–2 GB); 4 GB and 8 GB cost moreAbout 3 WNoNoNeeds 64-bit OS; SD cards slow and wear out
Rented server (VPS)$0; $6–12 a monthNone at homeNo (unless you rent a Mac)NoMany sites block data-centre addresses
Mac mini (M6)From $899About 4 WYesYes, sized by memoryCost; memory is chosen when you order

Prices and power figures are as of September 2026 and come from the sources listed at the end. For a sense of running cost: a box idling at 4 W uses about 35 kWh a year, one at 10 W about 88 kWh, and one at 20 W about 175 kWh. Multiply by your electricity rate.

An old laptop

An old laptop works if it can run a current 64-bit operating system and stay awake. OpenClaw's FAQ lists the downsides of running on a laptop: sleep and network drops disconnect it, and updates and reboots interrupt it. On Linux, closing the lid suspends the machine by default; you change that with the HandleLidSwitch setting in systemd's logind.conf. An old Mac laptop must be on macOS 13.5 or later. Wiping it and installing Ubuntu is often the simpler route if you don't need iMessage.

An office-surplus desktop or mini PC

Small ex-office PCs suit most people who don't need a Mac. I run my own agents on five computers: two old mini PCs, two office-surplus Dell OptiPlex desktops and a newer mini PC. ServeTheHome measured a Dell OptiPlex 3070 Micro at just over 10 W idle and expects this class of 1-litre office PC to idle at 9–12 W. That's more than a Mac mini, but you may already own one, and they're sold used. They're fine for OpenClaw with cloud models. For local models they only have the processor, and OpenClaw's managed setup stops recommending models above 9 billion parameters without graphics acceleration.

A Raspberry Pi

OpenClaw has an official Raspberry Pi guide and calls the Pi 5 the best choice and a 4 GB Pi 4 the "sweet spot for most users". The minimum is 1 GB of RAM and a 64-bit OS; 2 GB or less needs swap space added. The guide recommends a USB SSD over an SD card, and says not to run local models on a Pi at all because even small ones are too slow. Prices have risen: on 1 April 2026 Raspberry Pi added $25 to 4 GB boards and $50 to 8 GB boards because of memory costs, while holding 1 GB and 2 GB Pi 4 and Pi 5 boards at $35–65. You'll also need a power supply, storage and a case. Most features work on ARM; a few optional tools that skills install may lack ARM versions.

A rented server (VPS)

A VPS gives you a machine that's always on without anything running at home. OpenClaw's DigitalOcean guide uses the 1 GB plan, which costs $6 a month; the 2 GB plan is $12. The docs also point to Oracle Cloud's Always Free ARM tier, while warning that signup "can be finicky". They suggest a VPS if your agent has dropped off before. The tradeoffs: the server often has no screen, so browser tasks run headless; your files aren't on it; and the docs warn that many websites block data-centre addresses, so browsing from a home computer "often works better". Keep the Gateway private and reach it through an SSH tunnel or Tailscale, as the docs describe.

Keeping it running

Whatever you choose, the machine has to stay on and restart by itself. On a Mac mini, open System Settings, then Energy, and turn on "Prevent automatic sleeping when the display is off" and "Start up automatically after a power failure". On Linux, OpenClaw's installer sets the Gateway up as a systemd user service; on a headless machine, also run sudo loginctl enable-linger "$(whoami)" so it keeps running when you log out. OpenClaw's FAQ recommends a dedicated machine for reliability and isolation, with your everyday laptop paired as a node if you want the agent to use it.

I built ibara, free, open-source software that gives an AI agent a computer of its own, used through MCP tools. It works with any agent that can use MCP tools and is built for Omarchy (Linux) today, with Mac and Windows in progress. If you're still choosing an agent, Hermes Agent vs OpenClaw compares the two.

Questions

Can I run OpenClaw on an old laptop?

Yes, if it can run a 64-bit operating system with a supported Node version and has at least 1 GB of free RAM (2 GB or more is better). Old Macs must be on macOS 13.5 or later. Set it to stay awake with the lid closed, and expect it to go offline whenever it sleeps, restarts or loses Wi-Fi.

What are OpenClaw's minimum hardware requirements?

As of September 2026, OpenClaw's docs give an absolute minimum of 1 CPU core, 1 GB of RAM and about 500 MB of disk, and recommend 2 GB or more of RAM. It runs on macOS, Linux and Windows and needs Node 24.16+ or 26.1+. Running AI models locally needs far more: at least 8 GiB for OpenClaw's smallest managed local model.

Does OpenClaw work on a Raspberry Pi?

Yes. OpenClaw has an official guide. A Pi 5 or a Pi 4 with 4 GB works best; a 2 GB board works with swap space added, and 1 GB is tight. You need a 64-bit OS, and you should use cloud models, not local ones.

Do I need a Mac for OpenClaw's iMessage support?

Yes, but any Mac signed in to Messages will do, not only a Mac mini. OpenClaw itself can run on another computer and reach the Mac over SSH. Some iMessage features, such as reactions and threaded replies, need System Integrity Protection turned off on that Mac.

How much memory should a Mac mini for OpenClaw have?

16 GB, the base amount, is plenty if you use cloud models. For local models, memory sets the size of model you can run: OpenClaw's managed setup wants 24 GB for a 12-billion-parameter model and 32 GB for its largest options. Apple sells memory as an option you choose when you order, so decide up front.

Is a VPS better than a Mac mini for OpenClaw?

For staying online cheaply, often yes: a 1 GB DigitalOcean server costs $6 a month and stays on without anything running at home. A home computer is better for browsing sites that block data-centre addresses, for using local files, and for iMessage and local models, which a basic VPS can't do.

Sources