Best Mac Mini for OpenClaw: Which One to Buy, and How to Keep It On (2026)

Do you need a Mac mini for OpenClaw?

No. A Mac mini is one option among several. (People also write "Mac mini Open Claw": it is the same program.) OpenClaw's own hosting FAQ answers "Do I have to buy a Mac Mini to install this?" with "No." It runs on macOS, on Linux, the free operating system most servers run, and on Windows through WSL2, Microsoft's way of running Linux inside Windows.

People buy one because an assistant has to stay switched on. A laptop leaves the house with you, and your main computer needs restarting. Geoff Fagien, a video creator who covers Macs, described the appeal in October 2026: a Mac mini "can just sit there" as "the little always on server".

Whether it has to be that computer depends on how OpenClaw works. It has two parts that people mix up:

• The gateway is the program that stays on. It holds your passwords and API keys, the secret codes a program uses to sign in to a paid service. It receives your messages and runs jobs at set times. It also calls tools, the actions your agent can take, such as searching the web or sending an email.
• The model is the AI that does the thinking. Claude and GPT are models. A cloud model runs at the company that makes it: the gateway sends it each request over the internet, and your Mac does very little. A local model runs on your own machine, and then memory decides what fits.

With a cloud model, the new chip hardly matters. Fagien concludes that the M6's AI hardware "may not matter that much" for this work, and that many of the same jobs could run on an older Mac mini.

Three reasons make a Mac mini the better host:

• iMessage needs a Mac. OpenClaw's iMessage channel works only through a Mac signed into Messages.
• Apple's apps need a Mac too. OpenClaw's add-ons for Apple Notes and Reminders run only on macOS. Sean Matthew, a creator who runs an agent on a Mac mini, gave the agent its own Apple ID, the account Apple's services use, and added it to an iCloud family, Apple's way of sharing between people. The agent now shares Matthew's calendar and reminders. Matthew can send it files from an iPhone with AirDrop, Apple's wireless transfer.
• A local AI runs well on a Mac. If your messages and files must never leave the house, the model has to run on your own machine. A Mac can load a large model because its processor and graphics chip share one memory, where a PC needs an expensive graphics card.

Two other reasons call for a computer at home, not for a Mac in particular:

• A home internet address opens more websites. Many sites block traffic from data centres, where rented servers sit. OpenClaw's hosting advice says "local browsing often works better". Peter Steinberger, who created OpenClaw, said on the Lex Fridman podcast in February 2026 that a residential IP, meaning a home internet address, "makes a lot of things simpler." This matters when your agent browses the web for you.
• A separate computer protects your accounts. An agent that can run commands and read files is safer on a machine that holds none of your personal accounts. Limiting what it sees helps too: I never wanted to give OpenClaw my personal information, only data cut up for each job and worth little on its own.

Against buying, OpenClaw's docs make a small Linux server their default "for an always-on Gateway and low cost". Steinberger's own advice, in the same interview, is to reuse a computer you are replacing: "use the old one as your server instead of buying a standalone Mac Mini."

Owners disagree on whether a computer at home beats a rented server:

• A rented server is the cheap way to try it. Oliur, a creator with a spare Mac mini, told viewers in February 2026 not to buy one just to try OpenClaw, in a video that walks through a hosting company's one-click setup. After a power cut, Oliur says, a Mac at home can need you in front of it, while a server keeps running. Superbash, who makes AI videos on the BoxminingAI channel, likes the server's reinstall button: when an agent makes a mess, you start again from nothing.
• A computer at home is easier to keep private. Syed Humair, who runs an agent on a Mac Studio, Apple's larger desktop, points out that a rented server has a public address, one anyone on the internet can try to reach. So "one simple mistake, and your machine is exposed". Your home router, the box that connects your home to the internet, blocks incoming traffic by default. Humair calls that Mac Studio overkill and recommends an old laptop.
• A Mac at home is easier to live with. Alex Finn, a YouTuber who runs OpenClaw on a Mac Studio, likes being able to AirDrop a file to the agent from a phone. Finn also claims, without proof, that server companies paid creators to recommend them.

For most readers, the choice comes down to this (prices read on 4 November 2026):

The best Mac mini for OpenClaw at each price

This M6 replaced the Mac mini whose price most guides still quote. The base M4 started at $599 and sold out at Apple in spring 2026. In June its entry price rose to $799, as memory chips got more expensive, Macworld reported. On 25 August Apple announced the M6 Mac mini, on sale since 22 September.

Apple markets it as "the leading desktop for always-on agentic computing". That means a computer AI agents work on around the clock.

In the table, "custom" means you choose the memory in Apple's online configurator, and the other builds are ready-made. The M5 Pro is the larger chip, despite its lower number.

Sources: Apple's store and its configurator (M6: memory +$200 for 24GB, +$400 for 32GB; storage +$200 for 512GB, +$500 for 1TB. M5 Pro: memory +$600 for 48GB, +$1,000 for 64GB), before tax, 4 November 2026. Education price for the base M6: $799, from Apple's announcement of 22 September.

Decide the memory first. The memory is fixed when you order, so you can't upgrade it if you change your mind later.

Storage matters less. OpenClaw itself is small, and each local model is a file of 3 to 17 GB. So 256GB holds a cloud setup and a few local models, and 512GB leaves room for your files. The M6 has five USB-C ports, so an external SSD, a fast plug-in drive, can add space later. Fagien bought the 256GB M6 on purpose and keeps the models on external drives.

At $1,299, the two builds are not equal for a local model. The 32GB one, with 256GB of storage, runs the model OpenClaw prefers, which took 26 to 28 GB on a real job. That is more than the 24GB one holds, even with its 512GB of storage. The model's file is 16.5 GB, so it fits on 256GB. Take the 32GB one if a local model is the reason you're buying.

The M5 Pro is worth paying for only to get more memory or faster local answers. Memory bandwidth is how fast a chip reads its memory, and once a model fits, it largely sets how fast the words come out. Apple's specifications list 153GB/s for the 16GB M6, 170GB/s for the 24GB and 32GB M6, and 307GB/s for the M5 Pro. With a cloud model, none of this changes anything.

An older Mac works too. OpenClaw's Mac app requires macOS 26.2 or later, and Apple's lists for macOS 26 and macOS 27 both start at the Mac mini M1 from 2020. With a cloud model, an M1 or M2 with 16GB does the job as well as a new one.

Older still gets harder. The 2018 Intel Mac mini stops at an older macOS, so it can run the gateway from the Terminal, the Mac's command window, but not the app. Sean Matthew first tried a 2015 iMac, spent two hours trying to install the current Node.js (the software OpenClaw runs on), and bought a Mac mini instead.

Buying from anyone but Apple takes one extra check. Apple's refurbished store listed no Mac mini on 4 November 2026, so a cheaper one means another seller. Brian Johnson, a security consultant who hosts the 7 Minute Security podcast, bought a sealed Mac mini from an Apple-authorized seller on Amazon, after Apple quoted a four-week delivery. At first boot, the Mac said it was managed by another company.

Its serial number was tied to that company's device management, a system that lets an organisation control its Macs remotely. The seller removed the link within the hour, and Johnson erased the Mac and reinstalled macOS. Your agent's keys shouldn't sit on a Mac someone else can manage. If your first screen says a company manages the Mac, have the seller remove the link, then erase the Mac and reinstall macOS.

Skip anyone charging for the install. During that spring shortage, Alex Ziskind, a developer and hardware reviewer, found a base Mac mini listed nearby at $1,500 with OpenClaw pre-installed, and someone asking almost $900 to install it at your home. OpenClaw itself is free.

Which Mac mini for OpenClaw: how much memory you need

How much memory you need depends on where the model runs. The gateway is small on any Mac. The model is large only when the Mac itself runs it.

The gateway takes about 1.3 GB. To see what the agent itself costs, this page's test started OpenClaw on 4 November 2026 with nothing connected and watched its memory for five minutes. The number of processor cores barely changed the result: about 1.2 GB on one core, 1.3 GB on sixteen. While starting, it briefly reached 1.6 and 1.9 GB, in that order.

The same test on an earlier version, nine days before, read about 1.0 and 1.1 GB, so the figure is growing.

Here is what that leaves on a 16GB Mac. macOS takes 3 to 4 GB and each browser your agent opens takes 200 to 400 MB, according to SF AI Labs, a consultancy that runs OpenClaw on Mac minis for its clients. So 16 − 4 − 1.3 leaves about 10 GB, room for about 25 browsers. With a cloud model, that is a wide margin.

Test run for this page on 4 November 2026: official image ghcr.io/openclaw/openclaw, version 2026.10.10, in a Linux container, memory read from the container's control files, one run per setting. No Mac mini was tested for this page: the Mac mini results below come from the owners and reviewers named with them. The Mac app runs the same program; its exact use on macOS may differ.

Memory also fills up with what you forgot to close, such as windows and sessions left open. On a Mac you already own, open Activity Monitor to see what is still running before you decide you need a bigger Mac.

A local model needs far more. Its size is counted in parameters, the numbers it learned in training. In model names, B means billions of them. It also needs memory for its context window, everything it holds in mind while it answers. That context is counted in tokens, which are pieces of words.

OpenClaw can install a model for you through its llama.cpp plugin (llama.cpp is a widely used program for running models). The plugin's table gives the minimum memory it wants before offering each model. Qwen, Gemma and Muse are families of AI models you can download:

Source: OpenClaw docs, llama.cpp plugin, "Model recommendations", read 4 November 2026. Each uses a context of 65,536 tokens and can use tools. The docs call these "selection floors, not guaranteed fit or speed".

These minimums decide what setup offers, not what can physically load. Each one leaves room for a conversation of about 64,000 tokens, the size OpenClaw plans for. The table writes GiB and Apple writes GB, but for memory they are the same size: a Mac sold as 64GB holds exactly 64 GiB (read on 27 October 2026).

The 32GB Mac mini meets the top minimum with no margin. Setup also sets memory aside for macOS and checks what is in use, so on a busy Mac it may offer a smaller model. If it does, close your other apps and run setup again. If it still offers less, run the model in LM Studio, an app for running models that OpenClaw works with. The owners below loaded their models with such apps, not through OpenClaw's setup.

Their tests on the new Mac mini show what each size does:

• A local model on 16GB failed on agent work. Bart Slodyczka, who builds AI tools and reviews them, ran Gemma 4 12B on a 16GB M6 in September 2026, below its minimum in the table. It read invoices well. But its conversation was capped at 19,500 tokens, less than a third of the 64,000 above, and web-search results didn't fit. A file-sorting job then failed, with the model repeating the same step.
• 24GB runs a smaller model comfortably. The host of Better Creating, a YouTube channel on apps and creative work, used a 14B model on a 24GB M5 Pro, calling it "a safe bet". A writing job with personal files loaded took 22 GB. The host says the 27B model "technically should fit" but pushes the machine harder, and still calls the 24GB M6 the sweet spot for experimenting.
• 32GB runs OpenClaw's first choice, slowly. On a short test, Alex Ziskind's 32GB M6 ran Qwen3.8 27B at 23 GB of memory and about 10 tokens a second, a slow reading pace. On a real coding job for Execute Automation, a YouTube channel for developers, it took 26 to 28 GB and finished in 42 minutes. NightShift Tech, an owner reviewing after two weeks of daily use, reached about 21 tokens a second with Splash, a faster app for running models.
• 32GB still has a ceiling. Execute Automation's first try, with the larger Qwen3.6 35B, hit the memory limit after 8 minutes and the job failed.

That is the reason to take 32GB over 24GB for a local model. A short test fits in 24GB. Real work on OpenClaw's first choice took 26 to 28 GB, and memory can't be added later.

OpenClaw also sends the model a lot before your message: its instructions and the description of every tool. That came to about 16,000 tokens on a fresh install of version 2026.10.10. It was measured on 4 November 2026 by pointing OpenClaw at a fake AI server that only records what it receives, so no AI company was called.

A local model has to read all of it first. Andres Vidoza, a developer who logged a first setup on an M4 with 16GB, timed a small 7B model: 1.6 seconds to answer "hi" directly, 24 seconds through OpenClaw.

OpenClaw's FAQ is blunter: "Is a local model OK for casual chats? Usually no." It says small models forget the start of a long conversation and lack the safety filters a provider runs. It adds that they are easier to trick with instructions hidden in what they read, such as an email or a web page.

Owners who hoped for free AI found the same limit. Jeff Humble, a designer who co-founded The Fountain Institute, a design school, bought an entry-level M4 in early 2026 to run everything locally and pay no AI bills. The models were "about twice as slow as predicted and basically unusable". Humble switched to Haiku, Claude's cheap model, and spent $3.14 in an hour, mostly on troubleshooting.

A 32GB machine changes the picture for simple jobs. David Sparks, co-host of the Mac Power Users podcast, bought a 32GB M6 and summed it up on 25 October 2026: its models "are not gonna match what you get with an Opus frontier model" (Opus is one of Anthropic's most capable Claude models), but "for a lot of the donkey work… it is there now".

What the gateway, the model and the tools each keep on your machine is laid out in what self-hosted AI agents keep at home.

A Mac mini or something cheaper

If you need neither iMessage, Apple's apps nor a local model, the same agent runs on cheaper hardware. Each option has its own catch:

Sources: DigitalOcean, the hosting company these server prices come from, MacStadium, PiShop US for the Raspberry Pi, Apple's store; all read 4 November 2026.

The $6 server is too small on its own, and the $12 one isn't. In this page's test, the gateway limited to 1 GB of memory was shut down seconds after it started. Limited to 2 GB, it ran for the whole five minutes.

Swap, disk space a system uses as overflow memory, can make up the difference. An earlier version ran on 1 GB once 1 GB of swap was added, though with no room to spare. OpenClaw's DigitalOcean guide tells owners of the $6 server to add 2 GB. Take the $12 server if you can.

A Raspberry Pi is a credit-card-sized computer, and it is no longer the $35 machine of its reputation. Memory prices pushed the 4GB Pi 5 from $70 to $110 in two steps in 2026 (February and April), though OpenClaw's Pi guide still quotes "$35-80". That guide rates the Pi 5 with 4 or 8GB "Best" and the 2GB Pi 4 "OK" with swap.

Owners do run it there. Gary Sims, who explains technology on the Gary Explains channel, runs OpenClaw on a Pi 5 with a fast SSD, on a ChatGPT Plus subscription. PJ Bell, a builder who filmed a full install, set it to start with Linux, so it starts with the Pi and needs no login. Lighter agents than OpenClaw also suit a Pi or a small server, and they are compared in OpenClaw alternatives.

An old laptop is the cheapest box at home, and the one that copes best with a power cut. Its battery carries it through a short one, where a Mac mini goes down unless it sits on a separate battery unit. Wipe it and install Ubuntu, the Linux system OpenClaw's FAQ calls "the best-tested Linux install path". A coding agent can walk you through the install.

On Linux, the gateway runs as a background service. One command, sudo loginctl enable-linger $(whoami), keeps it running with nobody logged in. OpenClaw's guide to running the gateway calls this "the supported way to keep the Gateway running without a login session", the very thing a Mac lacks.

One laptop setting needs changing. Closing the lid suspends a Linux laptop, according to the manual of systemd, the program that runs Linux's background services. Ask a coding agent on the laptop: "Make this laptop stay on when the lid is closed, even on mains power." On Linux, the laptop gives up iMessage, and few old laptops have the memory for a local model.

An old Mac laptop has a problem of its own, even plugged in. In a bug report on OpenClaw's GitHub page, a Mac laptop on mains power, with sleep turned off, still dropped into maintenance sleep, the short naps macOS takes for background chores. Over about 44 hours, the gateway froze 87 times, and its connections to Feishu, a workplace chat app, dropped each time. So test a Mac laptop for a day before you rely on it.

The Mac mini's own electricity is not the deciding cost. Apple measures the M6 at 4 watts idle and 70 at most. Four watts all year is 35 kilowatt-hours, about $6 at the US average residential price of 18.16 cents in August 2026, according to the US Energy Information Administration. Flat out all year, it would be $111. The purchase price counts for more: $899 buys 75 months of the $12 server.

Set up your OpenClaw Mac mini so the agent comes back after a restart

A Mac mini only earns its place if your agent answers after a power cut, an update or a crash. That takes more than switching off sleep, because of how OpenClaw starts on a Mac.

OpenClaw runs its gateway as a LaunchAgent, a background job that macOS starts when a particular user logs in and restarts if it crashes. Apple's guide says such a job "executes only while that user is logged in". OpenClaw's gateway guide says it does not install the other kind of job, a LaunchDaemon, which starts with the Mac itself.

On a Mac, OpenClaw starts when the account it runs in logs in, not when the Mac starts, so a restart can leave your agent waiting at the login screen. A user reported it in March 2026 on OpenClaw's GitHub page: after a reboot "without logging in", the gateway gave "no response", then "Log in → gateway starts". That report was still open on 4 November.

Automatic login would solve it, except that Apple requires a manual login when FileVault, the Mac's disk encryption, is on. So here is what a power cut leaves you with, once the steps below are done:

• With FileVault on, you have to step in. The Mac starts by itself and stops before macOS loads. From your home network, or through a VPN into it (an app that links your devices over a private network), you unlock the disk and, if your agent doesn't answer, log in to its account. Until then, your agent stays offline.
• With FileVault off and automatic login, nothing is left to do. The Mac starts, logs in to your agent's account, and your agent starts. Anyone who takes that Mac can read its disk, your API keys included.
• With a UPS, a short cut changes nothing. A UPS is a battery unit that keeps the Mac running through a short power cut.

The setup uses two macOS accounts. The admin account, the one you create when you first switch the Mac on, makes every change to the system. A standard account, which can't install software for the whole Mac or change its system settings, runs your agent. So a tricked agent can't do either. Lee and Eubanks both run OpenClaw this way.

A Mac mini comes with no screen, keyboard or mouse, and the first setup needs them, so borrow a set for it.

The steps use the Terminal for a few commands, which call the agent's account agent and the Mac mac-mini.local. Replace those with yours: System Settings, General, Sharing shows the Mac's address. When a command fails, ask a coding agent what the error means. Jeff Humble got through the install that way, with Claude Code diagnosing each failure.

In the admin account

1. Create the agent's account. In System Settings, open Users & Groups, click Add User and choose Standard. Sign nothing personal into it. Keven Builds, a creator who runs four Mac minis, gives each one a throwaway email, and Sean Matthew's agent has its own address through AgentMail, an email service built for agents.

2. Stop the Mac from sleeping and make it start after a power cut. In System Settings, open Energy and turn on "Prevent automatic sleeping when the display is off", "Wake for network access" and "Start up automatically after a power failure", as Apple's Energy settings page lists them. Or set them from the Terminal with the first line below. Either way, run the second line to check:

   Copy

   sudo pmset -a sleep 0 autorestart 1 womp 1
   pmset -g | grep -E "^ *(sleep|autorestart|womp) "
   

   The check should show sleep 0, autorestart 1 and womp 1 (womp is Wake for network access). The sleep line may add "(sleep prevented by …)" with app names. Those are apps keeping the Mac awake right now, and you can ignore them. If the sleep number isn't 0, run the first line again.

3. Switch on remote access. Remote Login is the Mac's SSH access, the standard encrypted way to log in to another computer. In System Settings, open General, then Sharing, and switch on Remote Login and Screen Sharing. Under Remote Login's "Allow access for", choose "Only these users" with the admin and the agent accounts.

4. Let the agent's account unlock the disk. FileVault's switch is in System Settings, Privacy & Security. Since macOS 26, a Mac with an M-series chip can be unlocked remotely: Apple's FileVault guide says it "can be unlocked over ssh after a restart if Remote Login is turned on and a network connection is available".

   Only an account allowed to unlock the disk can do it. Add the agent's account with sudo fdesetup add -usertoadd agent, as Eubanks does for each new account. sudo fdesetup list shows the accounts allowed.

   After the unlock, SSH disconnects while macOS finishes starting. Apple doesn't say whether the account you unlocked with is also logged in, so your own test shows whether yours needs a login too. Four more things are worth knowing:

   • The unlock works from your home network. Tailscale, a popular VPN app, and others like it only start after the unlock. From outside, you need a VPN on your router or on another device at home.
   • A remote-control box may reach it from anywhere. One host of Mac Power Users plugged a GL.iNet Comet Pro into a Mac mini as a second way in, and reaches it through Tailscale. The box acts as a screen and keyboard on its own network connection, and its maker says it works before the operating system has started. This page hasn't tested it on a FileVault unlock.
   • Wired is the safer connection. Jeff Geerling, who writes about home servers and unlocks Macs this way, found at first that only Ethernet worked before the unlock. Wi-Fi worked for Geerling after macOS 26.5 shipped in May 2026, but plugging in removes the doubt.
   • A planned restart can skip the unlock. sudo fdesetup authrestart restarts the Mac and skips the unlock once. Apple's manual for that command warns that FileVault protection is reduced until the Mac has restarted, because a copy of the unlock key is kept in memory.

   If you chose FileVault off, turn it off there, then in Users & Groups set "Automatically log in as" to the agent's account.

5. Control updates. Under General, then Software Update, Apple can download macOS updates and install them "later that night", a restart you didn't plan. Turn off "Install macOS updates" on your agent's Mac. Install them yourself when you're at home, restarting with sudo fdesetup authrestart.

   OpenClaw's own automatic updates are off by default, so its version changes only when you ask. An OpenClaw update has broken my install more times than I can count. So update it when you have time to repair the setup, and weigh that time against what your agent gives you back.

In the agent's account

Log out, then sign in to the agent's account at the Mac or through Screen Sharing.

6. Install OpenClaw with the Mac app. Its setup needs no Terminal and no admin rights, according to OpenClaw's Gateway on macOS page. Download it from OpenClaw's macOS page, copy it to Applications while logged in as admin, open it in the agent's account, and choose This Mac.

   Then turn on "Keep OpenClaw running when the app is closed" next to "Launch at login" in its device settings. That switch is off on a fresh install, and without it the gateway stops when the app quits.

   The Terminal is the other route, for a Mac on an older macOS. First, in the admin account, run xcode-select --install: it installs Git, a program OpenClaw's installer needs. Then, in the agent's account, use OpenClaw's other installer, which needs no admin rights, unlike the usual one:

   Copy

   curl -fsSL https://openclaw.ai/install-cli.sh | bash
   ~/.openclaw/bin/openclaw onboard --install-daemon
   

   The second line connects your AI provider and sets OpenClaw to start when this account logs in. Add a chat app such as Telegram afterwards with openclaw channels add, as the onboarding guide suggests.

7. Keep the gateway private. By default it accepts connections only from the Mac itself, and OpenClaw's remote access page advises leaving it that way. You still want to reach it from your laptop, because it serves the Control UI, OpenClaw's web dashboard, where you chat with your agent and change its settings.

   Tailscale's Serve feature or an SSH tunnel does that. A tunnel carries the gateway's port, its numbered entry point (18789), over an encrypted link. From your laptop:

   Copy

   ssh -N -L 18789:127.0.0.1:18789 agent@mac-mini.local
   

   The window stays blank and must stay open while you use the dashboard. Open http://127.0.0.1:18789 in the laptop's browser: that address means "this computer", and the tunnel carries it to the Mac. Anyone holding that dashboard's login can do what you can do with your agent, so keep it private. What it allows is compared in Hermes vs OpenClaw, Hermes Agent being an open-source rival from Nous Research.

Some owners write their own LaunchDaemon so the gateway starts without a login. The user who opened that GitHub report said it started about 37 seconds after boot, with nobody touching the Mac. Skip this route unless you can maintain it yourself: if such a job uses OpenClaw's service name, OpenClaw refuses to install or repair its own. Nothing starts before a FileVault disk is unlocked anyway, and iMessage still needs the Messages app in a logged-in account.

Test the restart before you rely on it

Guides that promise a fully automatic recovery leave out the login, so cause one real power cut while you are at home, with a laptop on the same network.

1. Unplug the Mac for ten seconds, plug it back in, and give it a minute to start.
2. If FileVault is on, unlock the disk with the agent's account: run ssh agent@mac-mini.local and type that account's password. The connection closes while macOS finishes starting, as Apple's manual says it will.
3. Send a harmless message through the chat app you use, such as "reply with today's date", and wait for the answer.
4. If no answer comes, the agent's account isn't logged in yet. On a Mac laptop, open Finder, choose Go, then Connect to Server, and type vnc://mac-mini.local. From Windows or a phone, use any VNC app, the same kind of screen sharing. Sign in to the agent's account and send your message again.
5. If there's still no answer, check the gateway in that account. The OpenClaw app shows it in the menu bar. After a Terminal install, ~/.openclaw/bin/openclaw gateway status should show Runtime: running and a successful connectivity check.

Whatever you had to do by hand here, every long power cut will ask of you too. If that is too much, add a UPS, a VPN into your home network or, as a last resort, turn FileVault off.

One more thing to watch: macOS restarts a gateway that crashes, not one that hangs. A Mac mini M5 Pro owner reported in September 2026 a gateway that used all of a processor core and stopped answering until restarted. A fix for a likely cause has shipped since, but the report's trigger was never confirmed.

A scheduled check catches a hang. In the agent's account, paste this one line into the Terminal. macOS may ask you to allow the Terminal to make the change. It adds a job that asks the gateway for a sign of life every 10 minutes, and after 20 seconds without an answer, restarts it. This page hasn't run it on a Mac mini:

(crontab -l 2>/dev/null; echo '*/10 * * * * curl -fsS -m 20 http://127.0.0.1:18789/healthz >/dev/null || launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway') | crontab -

The check can only help while the agent's account is logged in, since the gateway's LaunchAgent exists only then.

This test can't decide the purchase, since it comes after it. Your list decides that: iMessage, Apple's apps or a local model mean a Mac mini, and nothing on the list means one of the cheaper options above. The test tells you what an outage will ask of you.

iMessage: the feature that needs a Mac

iMessage is the strongest reason to give OpenClaw a Mac, and it comes in two levels. OpenClaw reaches it through imsg, a small program that drives the Messages app.

• Basic mode changes nothing on your Mac. Your agent sends and receives text, photos and files.
• Private API mode adds everything else. It uses parts of Messages that Apple keeps private. That means replies in a thread, reactions, edits, unsend, message effects, polls and group management. It needs System Integrity Protection off, the macOS safeguard that stops anything from modifying system software. It also needs a second protection relaxed, the check on what can load inside Apple's apps. imsg then places a small helper program inside Messages.

OpenClaw and imsg disagree on that second level. OpenClaw's iMessage page calls private API mode "strongly encouraged", and in basic mode its plugin hides those extra actions from your agent. The guide from the imsg project tells most readers "don't", because macOS limits make these features "unstable in production".

To connect basic mode, follow these steps from OpenClaw's setup page. This page hasn't run them on a Mac mini. If one stalls, ask a coding agent: "Connect iMessage to OpenClaw on this Mac in basic mode, following docs.openclaw.ai/channels/imessage/setup."

1. Give your agent its own address. Create an Apple ID for it, and sign in to Messages with that ID in the agent's account. The docs' deployment patterns recommend this, so your agent texts from its own address rather than yours.
2. Install OpenClaw's plugin. In the agent's account, run openclaw plugins install @openclaw/imessage.
3. Install imsg from the admin account. The usual way is brew install steipete/tap/imsg, and Homebrew, a program that installs other software, needs an admin account. Stephen Lee's setup installs such things from the admin account, then returns to the standard one.
4. Connect the channel. Back in the agent's account, run openclaw channels add --channel imessage.
5. Grant two permissions in that account. In System Settings, Privacy & Security, give Full Disk Access and Automation to OpenClaw when macOS asks.
6. Approve your first message. Text your agent from your phone. openclaw pairing list imessage then shows a code, and openclaw pairing approve imessage followed by that code lets you in. The code expires after an hour.

Private API mode adds a step after every restart, because macOS can reopen Messages without the helper program. Since version 0.19.1 of 22 October 2026, imsg launch --install-agent restores the helper at login and checks it every five minutes. Without that, run imsg launch yourself. Like your agent, the helper can only start once the agent's account has logged in.

Messages sent while your agent is down are not all lost. When the gateway runs on the Messages Mac, OpenClaw replays what it missed on restart, up to 500 messages and about two hours old. Older messages are dropped, so a long outage still costs you some.

iMessage always needs some Mac signed into Messages, but not necessarily a new one:

• A Mac you already own can do it. OpenClaw's hosting FAQ says so directly: "You need some macOS device signed into Messages - not necessarily a Mac mini, any Mac works." It still has to stay on.
• The gateway can stay on a server. It then works the Messages app on a Mac you keep signed in. A short script does this over SSH. The docs warn that sends can fail with a permission error while everything else looks fine, so test a real outgoing message first.
• The Mac can be rented. MacStadium rents an M4 Mac mini with 16GB in its data centre for $149 a month (4 November 2026), so six months cost what the base M6 does.

You may not need iMessage at all. Telegram, WhatsApp and Discord connect to OpenClaw with no Mac in the picture. I don't use iMessage with my agents: they message me on a private Discord server, and these days they send more reports than they hold conversations.

What a Mac mini OpenClaw setup costs to run

The running cost that matters most is the AI, and it is the same on any box. Part of it runs even when you send nothing.

OpenClaw's setup creates a check that wakes your agent every 30 minutes, listed as "Heartbeat (main)" and called a proactive check in its docs. The check is what lets your agent speak first: it reads a checklist you can edit and messages you if something needs attention. With it off, your agent still answers you and still runs the jobs you schedule.

Here is what it did in a test on OpenClaw 2026.10.10, on 4 November 2026, with the fake AI server:

• With no chat app connected, the check was skipped and the model was never called.
• Forced to run, it sent the model about 13,300 tokens each time, less than a first chat's 16,000 because it loads fewer instructions. With a chat app connected, it has somewhere to send its result, so expect the same, though the test didn't run that case.

With an Anthropic API key, billed per token, the sum runs like this:

• 48 checks a day × 30 days = 1,440 checks a month, × 13,300 tokens = about 19 million tokens.
• Anthropic says its newer models, Sonnet 5.5 and Haiku 5.5 among them, count about 30% more tokens for the same text. That makes about 25 million.
• OpenClaw asks Anthropic to keep the prompt, the instructions and tool descriptions sent each time, in its five-minute cache, a temporary store that makes a repeat read cheaper. The cache has expired by the next check, so every check pays the price of writing to it.
• On Claude Sonnet 5.5, Anthropic's mid-priced model, writing costs $2.50 per million tokens on its price list (4 November 2026). 25 × $2.50 = about $62 a month, before you send a single message.
• On Claude Haiku 5.5, a small model at $0.125 per million, the same sum gives about $3.10.

These are estimates, not bills, because the prompt grows as the conversation does. You have three ways to lower them:

• Turn the check off. openclaw config set agents.defaults.heartbeat.every '"0m"' --strict-json stopped it in the test. The docs' own command, openclaw automations disable, was refused on that version.
• Point it at a cheap model. openclaw config set agents.defaults.heartbeat.model '"anthropic/claude-haiku-5-5"' --strict-json names the model for the check alone. In the test, the check then used the model named. On Haiku 5.5, that is the $3.10 above.
• Run it less often. The first command with '"55m"' moved it to every 55 minutes. That stays inside OpenClaw's one-hour cache, so each check pays the repeat-read price, $0.10 per million on Sonnet 5.5, or about $1.40 a month. To switch that cache on, ask a coding agent to set OpenClaw's cacheRetention to long. The docs warn that a check can still miss the cache.

Set a spending cap too, at your AI provider when it offers one. I give my agents one regularly. I've never needed it, and it costs nothing.

A subscription login, such as the ChatGPT sign-in that a fresh OpenClaw setup offers, works differently. The check then uses your plan's allowance instead of costing per token, and your agent stalls when that allowance runs out until the plan's limit resets.

Your action plan

Every action in this guide, in the order of the steps, to tick as you go.