How to Run Claude Code 24/7 on a Mac Mini: 7 Steps and the Janitor Jobs I Added After It Choked

$ launchctl list | grep -cE 'com\.(chris|shiplean)'
104

Step 1: Pick the job before you pick the machine

Write down one job that should run while your computer is off.

“But have a use case. Don’t just buy it to buy it.”

I bought the Mac mini for OpenClaw (an always-on AI agent app I ran before) when it first came out. Then Hermes, another agent app, ran on it. Then both got nuked, and I never really found a use case for either after that. On October 6 I retired both.

“Grok Bot became the first agent where it was like, okay, this is different.” Now I’m working with my ChatGPT dot, which OpenAI’s help center calls “an always-on agent in ChatGPT.” It’s always changing, man.

What stuck is the station. When I edit videos, all of that happens on the Mac mini, so it’s an app I can reach 24/7.

If you do buy one, you don’t need the latest. My first guess was a hundred or two hundred on eBay. A search of used listings on October 23 showed asking prices closer to $250 to $325 for an M1 like mine. Those move, so check before you buy.

Fill this in before you spend anything:

Job: [the thing you'd run overnight, e.g. "edit tomorrow's video"]
How often: [every night / every Friday]
What I check in the morning: [the file, report or message it leaves]

Applicable for: anyone who already runs one long Claude Code job and has to babysit the laptop for it.

Step 2: Keep your laptop as home base, and set the other computer to never sleep

System Settings > Energy > Prevent automatic sleeping when the display is off

My main session is still on the MacBook Pro. The mini doesn’t replace it. It’s the machine that’s awake when the MacBook isn’t.

On the always-on computer, turn off sleep. Apple’s sleep guide gives the switch above for desktops. On a laptop it reads “Prevent automatic sleeping on power adapter when the display is off,” under Battery > Options.

Here is what my Mac mini reports:

$ pmset -g   (excerpt, October 23, 2026)
 autorestart          1
 disksleep            0
 sleep                0 (sleep prevented by Claude, caffeinate, powerd, mds_stores)

sleep 0 means it never sleeps. That’s the line you want.

Where it breaks: running caffeinate and calling it done. Apple’s manual page says its sleep block lasts “for the duration of the utility’s execution.” Reboot, and it’s gone. The Energy setting stays.

Step 3: Make it start back up after a power cut

System Settings > Energy > Start up when power is connected > Always (Mac mini from 2024 on, macOS 26.5 or later)

Apple’s power-on article says this setting also turns the Mac on “when power is restored after a power failure.” Older models like my M1 show a simpler switch on the same pane, “Start up automatically after a power failure,” per Apple’s Energy settings guide. That’s the autorestart 1 line in step 2.

Then check one more thing, because it decides whether your jobs come back. Apple’s automatic login article says manual login is required when FileVault is turned on.

So with FileVault on, a power cut leaves the Mac at the login screen. Your scheduled jobs start only after someone logs in. Decide which you’d rather have, and know which one you picked.

Step 4: Sync your files to it every five minutes

One sync job, every 300 seconds

This is the step that made it feel like one computer instead of two. “My Mac mini syncs up with my repo, so it’s always up to date, every five minutes.”

My folders live on GitHub, so the sync pulls the latest copy every five minutes. Its rule: if a file on the mini has edits that aren’t saved anywhere else, it refuses to overwrite it and logs the problem instead.

So I step away, the MacBook Pro is off, and I want to do something. The mini’s folders already match. As far as I’m concerned, it’s the same experience.

The one gap: I won’t see the chat I had open on the MacBook Pro. When I’m back on it, I ask the MacBook Pro to pull that chat up and resume there.

Yours doesn’t need GitHub. Whatever sync you already trust works, as long as both machines see the same folder.

Step 5: Make one word open a Claude session on it

One shortcut in your shell settings, then type the word

When I’m logged in on the Mac mini, I just type in “iris” and boom, it opens the Claude session immediately.

That’s the whole trick. If opening a session on the always-on machine takes six steps, you’ll go back to the laptop every time.

Here’s the shape of a shortcut like that, one line in your shell settings file (~/.zshrc on a Mac):

alias iris='cd ~/your-work-folder && claude'

Swap in your own folder and your own word. Open a new Terminal window, type the word, and Claude Code starts in that folder.

Step 6: Move every scheduled job onto the always-on machine

launchd, the Mac’s built-in scheduler, in ~/Library/LaunchAgents/

Everything has moved to the Mac mini. Everything saves there, and all the jobs run there.

Most of my 104 jobs are plain scripts. Only some of them call Claude Code headless, with claude -p (Claude Code’s headless mode: it runs a prompt with no chat window). A few run Codex or Gemini instead.

I use launchd because the in-app schedulers have limits for this:

One trap on headless jobs. Anthropic’s headless docs say --bare mode “doesn’t use your subscription login” and needs an API key. A job written that way bills the API, not your plan.

The launchd file that keeps Claude Code itself running, and how I steer it from my phone, are in my Claude Code Remote Control setup. I won’t repeat them here.

Step 7: Add the janitors before the night it chokes

Three timers and two nightly jobs, all in launchd

Back to that 50%:

“Today is Friday; by tomorrow morning I would have already been at 50% usage. That’s a big jump from right now: I’m at 15%, using it very lightly.”

The janitors came one scar at a time:

1. Tab janitor, every 30 minutes, since October 8. My robots open browser tabs and walk away. About 20 windows were open on the mini when it went in. It closes the tabs they leave sitting idle.
2. Memory sweep, 2:30 a.m., since October 16. The Mac mini has 8 GB, and that night swap hit 97% and a job timed out. It quits the apps and leftover processes that hog memory. This morning it took swap from 1,487 MB used to 851 MB.
3. Guard, every 15 minutes, since October 20. That day, test runs hung for hours. It kills test runs older than 2 hours and video renders older than 3, and it never touches Claude or Codex.

Two quieter ones run beside them: log cleanup at 2:30 a.m., and a health check three times a day that flags failed jobs.

How to tell they work: they’re boring. The guard has logged nothing but heartbeats since October 20. The tab janitor closed one or two tabs a handful of times last week.

Even now I’m a little inefficient. “We have an issue with our editing skill that I was tweaking last week and now we’re paying for it.” Janitors don’t fix a bad skill. They keep the mini from choking.

Check it worked: wake up to a review, not a mess

Every morning: read what ran overnight, then approve or skip

One of the overnight jobs fixes easy bugs. When one of my jobs breaks, the robot help desk picks it up at 3:30 a.m. Gemini classifies it, like IT support: what’s in the backlog, what’s pending?

“This is easy, hand it off to Codex. This is hard, hand it off to Claude.”

Codex prepares and tests the easy fixes. Claude Opus diagnoses the hard ones. Nothing ships on its own.

Then every morning I review them. My dot, the ChatGPT agent, messages me: this got fixed, do you want to fix it this way? I tap yes or I don’t. How I split work between the two tools in the daytime is in using Codex and Claude Code together.

It’s new. I built it on October 22, and last night’s run finished with nothing in the queue. That’s a good night.

Roby at Coding with Roby runs the same shape on his Mac mini: “It does work for me while I sleep, and I wake up to a queue of pull requests waiting for review.” Same rule. The machine works, a human ships.

Last check is the meter. Anthropic’s Claude Code plan page says the five-hour and weekly limits “are shared across Claude and Claude Code.”

My Claude usage meters, Max plan, September 5, 2026.

Every overnight Claude job draws from that same bar. If it jumped while you slept, open the logs before you open a new chat. How those meters reset is in Claude usage limits explained.

Have a use case, then let it run

You don’t need a new box to start. You need one job worth doing while you sleep, and a computer that stays awake for it.

Mine went from OpenClaw to Hermes to Grok Bot to my dot. It’s always changing. The setup underneath didn’t: never sleep, stay in sync, one word to open it, and janitors so it doesn’t choke.

My rule: have a use case. Don’t just buy it to buy it.

So tonight, plug in the laptop you already have, turn off sleep, and give it one job. Every other Claude guide for work, in order, lives on Claude guides.

FAQ

Do I need a Mac mini to run Claude Code 24/7?

No. Any computer that can stay plugged in and awake works, including an old laptop with the lid open. Buy a Mac mini only once you have a job that needs to run while your main computer is off.

Does caffeinate keep Claude Code running all night?

Only while caffeinate itself keeps running. Apple’s manual says its sleep block lasts for the duration of the command it wraps, so it does nothing after a reboot or a power cut. For a machine that should always be on, use System Settings > Energy instead.

Will my Mac turn back on after a power outage?

Only if you set it to. On a Mac mini from 2024 or later on macOS 26.5 or later, choose Energy > Start up when power is connected. Older Mac minis have Start up automatically after a power failure. If FileVault is on, the Mac then waits at the login screen, and your jobs don’t start until someone logs in.

Do Claude Code /loop tasks run forever?

No. Anthropic’s docs say recurring tasks expire 7 days after creation, firing one last time before they delete themselves. For anything longer, use the Desktop app’s scheduled tasks or the Mac’s own scheduler, launchd.

Does running Claude Code all night use up my plan?

Yes, it draws from the same pool as your daytime chats. Anthropic’s support page says the five-hour and weekly limits on Pro and Max are shared across Claude and Claude Code, so a stuck overnight job can eat the week you planned to use.

Where this comes from

The Energy settings, login rule and scheduler limits come from Apple’s and Anthropic’s docs, read October 23, 2026. The job list, schedules and power readout come off my Mac mini that same day.

The buy advice, the 50% night and the help desk are mine. I haven’t timed a power cut on my Mac mini, and the help desk only went live on October 22.

Published October 23, 2026. Checked that day against Apple’s and Anthropic’s docs, linked inline.