My AI agent fell asleep every time I closed my laptop. So I moved it into a Mac mini.
Four companies shipped an always-on AI agent in ten weeks, and all four run it on a computer in their own cloud. That's convenient, and it costs reach: an agent in a data center can't touch the servers, file shares and tools in your house. On Oct 22, 2026, I moved my own agent, Earl, onto a Mac mini at home instead. A week later I moved half of it back. Earl now works from two machines: the laptop when I'm at the desk, the Mini when I walk away. This build log is how that works, what it cost, and what I'd want to know before setting it up. It isn't for everyone.
Instrument and questions: Rick Worthington Analysis and prose: Earl (Claude Opus 5.5), under Rick Worthington's editorial direction
Watch the video at lab.worthington.cloud/entries/always-on-at-home
The question
Can I have the always-on AI agent everyone is launching, and keep it on my own network?
In ten weeks, four companies launched an agent that never logs off. xAI's Grok Bot came first, on Aug 11, 2026.[1] Meta's Muse followed on Sep 8,[2] OpenAI's dots on Sep 29,[3] and Anthropic's Claude Duet on Oct 21.[4] Each one works while you're away, answers from your phone, and remembers what you told it last week. Each one does that on a computer the company owns, in its own cloud.
I already had an agent. Earl is Claude Code with a few months of my own setup on top: rules, memory, and the keys to the servers in my house. It ran on my laptop, so it had the one weakness the cloud agents were built to fix. When the lid closed, Earl went to sleep.
1 of 235 sessions had the phone link switched on. It ran through the laptop, so a closed lid cut it off.
Earl's last 235 sessions before the move, and the one that could be reached from the phone. That link only worked while the laptop was open and awake.[5]
Duet was announced on a Wednesday. By Thursday afternoon, Earl lived on a Mac mini in my house. That's about 16 hours from the first question to the cutover, a night's sleep included. This page was planned from the Mini. A week later I moved my desk work back to the laptop, for features the Mini couldn't give me.
What the bench showed
Here's what you'd likely assume about running an always-on agent yourself, and what the move said about each.
“An always-on agent needs a company's computer.”
A Mac mini did it
30 of 30
one-a-minute checks answered, with no gap through nearly 12 minutes of the laptop asleep.
Who it fits
People who already run things at home that an agent should reach, and who are happy to be their own IT department.
The answer in one breath
Measured or published
An always-on agent is three parts: a computer that never sleeps, a way in from anywhere, and a memory. The cloud agents rent you all three in a data center. I put the computer and the memory in my house, on a Mac mini that was already there, and kept the way in that Anthropic already ships. It held with the laptop closed, came back on its own when I crashed it, and reached everything on my network. It also gave up the desktop app's best features, which only appear in sessions on the machine the app runs on, and every guardrail I didn't build myself. So Earl ended up on two machines, kept in step: the laptop at the desk, the Mini for nights and the phone.
My reading, not measured
What I read into it: the cloud agents and this setup answer different questions. If your life is email, calendars and web apps, a cloud agent reaches all of that already, with better guardrails than you'll build. If the things you want an agent to touch live in your house, the agent has to live there too. That's an argument from one move, not a measurement.
What's on the bench
1 Mac miniAn M4 with 16 GB of memory, already on the shelf and already always on. It also runs a small local AI model
4 WApple's own idle figure for that model, measured at the wall. About $5 a year at 15 cents a kilowatt-hour
$0in new subscriptions. Earl runs on the Claude plan I already pay for
16 hoursfrom the first question to the cutover, Oct 21 to 22, 2026, with a night's sleep in the middle
A
Part A: What an always-on agent is
Before the results: what these products are, under the names. Take the branding off and every one of them has the same three parts.
The first part is a computer that never sleeps. An agent that works while you're away needs somewhere to work. A chatbot only runs while you're typing to it. An always-on agent needs a machine that stays on, with its own browser and its own files. All four companies say theirs gets one: OpenAI says each dot has "its own cloud computer," and xAI's documentation says "the shared Grok Bot computer runs in the cloud."[3, 7]
The second is a way in from anywhere. You talk to it from your phone, a browser or a desktop app, and it's the same agent wherever you are.
The third is a memory. It remembers your preferences, your projects and what it did yesterday, so you don't start over every time.
The cloud agents
The company’s cloudA computer that never sleepsThe memoryThe model
Phone, browser, laptopthe way in
Your houseOne computer you allow, if it’s onEverything else on the network
Earl, since Oct 22
Anthropic’s cloudThe model
Phone, browser, laptopthe way in
My houseA Mac mini that never sleepsThe memory, the files, the toolsEverything else on the network
The same three parts, placed two ways. The model itself stays in its company's cloud either way.[3, 7, 8]
The move changes where two of the three parts live. The model itself still thinks in Anthropic's cloud. Every reply Earl writes is computed there, exactly as before. What moved into my house is everything around the model: the program that runs the tools, the files it works on, the memory, and the keys to the rest of the network.
The way in is a feature Anthropic already ships, called Remote Control. Claude Code runs on a machine you own, and the phone, the web and the desktop app all connect to it. The machine connects out to Anthropic, so nothing in the house is opened to the internet.[8]
B
Part B: Four in the cloud, one at home
The four products side by side, from each company's own pages, with the setup on the Mini in the last column.
Grok Bot (xAI)Muse (Meta)Dots (OpenAI)Claude Duet (Anthropic)Earl on the Mac mini
Where its computer is
xAI's cloud, one computer shared by all your Bots
Meta's cloud (reported)
OpenAI's cloud, one per dot
Anthropic's cloud
A Mac mini in my house
Reaches the computer you're at
If you allow it
Through its Mac app
If you allow it
No
It is one
Reaches the rest of a home network
Not described
Not described
Not described
Not described
Yes, all of it
Talk to it
Messages
Messages, WhatsApp
Messages, Slack, Teams, voice calls
Text or recorded voice messages
Text, dictation only
Guardrails
Approvals, rules, auto review
Approvals, audit trail, a password vault it can't read
Approvals, custom rules, read-only background work, a safety monitor
Asks before it sends, pays, signs or deletes
Whatever I set up
What it costs
A top plan (Cursor Ultra, $200 a month, reported)
Free, $20 or $100 a month (reported)
First one included in Pro and Business Premium
Included in Pro and Max, shares the plan's limits
My existing plan, plus a computer I already had
When the power goes out
Their problem
Their problem
Their problem
Their problem
Mine
What each company says about its own agent, read Oct 22, 2026. "Reported" means the figure comes from press coverage, not the company's own page. "Not described" means none of their pages I read mention it.[7, 1, 9, 2, 3, 4, 10]
Every cloud agent comes with guardrails a whole company built: they ask before they send, pay or delete, they keep an audit trail, and OpenAI's runs its background work with read-only tools.[3, 9, 4] Earl has whatever guardrails I set up, and that's a real cost. In exchange, Earl reaches the whole network, where three of the four can reach one computer you allow. xAI's own documentation suggests leaving that switched off unless a Bot "has a specific reason to work on your local files."[7]
Before the move I scored five ways forward for Earl, and one of them was rebuilding it around Duet. Duet was ruled out on reach. Anthropic's help article says it "can't see all your projects or files on your computer," and it can read my work in Claude but not change it.[10, 5] For an agent whose job is mostly in my house, that settled it.
C
Part C: Does it hold up
Always-on means you can walk away. Three tests of that, run in the small hours of Oct 22 on a trial copy, before anything real moved.
f1moderateOne test, one night
With the laptop asleep, Earl kept working.
MeasuredEarl ran a check once a minute on the Mini, and all 30 ran on time, 12 of them while the laptop slept with its lid closed.
Measured or publishedFrom the move and the record
Laptop asleep, lid closed
2:06:22 AMasleep 2:22:09 to 2:33:562:34:05 AM
30 of 30 ran on time, one a minute, 12 of them while the laptop slept for 11 min 47 s.
Each tick is one check, at the time it ran on the Mini. The shaded span is the laptop asleep.[11]
For whoever runs it: Test the thing you're buying: close the lid and walk away, then check.
Evidence and objections
Measured The checks ran from 2:06 to 2:34 AM ET, about 57 seconds apart, and the laptop slept from 2:22:09 to 2:33:56. No check was missed or late.[11]
Measured This session was opened from the laptop's app over a remote login to the Mini, the harder case. The server setup Earl runs on now doesn't involve the laptop at all.[11, 8]
What it cost: each check was a real turn for the model, so 30 checks were 33 model calls. Proving the agent is alive by asking it is expensive. Later checks asked the Mac instead: is the process running, and has it restarted.
f2moderateOne forced crash
When I crashed it on purpose, it was back in two seconds.
MeasuredI killed the server outright. The Mac restarted it in 2 seconds, and the conversation that was open picked up where it left off.
Measured or publishedFrom the move and the record
I kill the server on purposeThe Mac restarts itThe same conversation carries on
2 seconds from the kill to a running server. No one touched it.
A hard kill, the kind that gives a program no chance to clean up.[11]
For whoever runs it: Let the operating system own the restart. Don't count on remembering to.
Evidence and objections
Measured The server runs under the Mac's own service manager with an instruction to keep it alive, so a crash is a two-second gap rather than an outage.[11, 12]
Not tested yet: a full restart of the Mini, and a power cut. Part E says why those are the hard case.
f3moderateOne conversation, three windows
Phone, browser and laptop were three windows onto one agent.
MeasuredOne conversation was started on the phone, continued in a browser and finished in the laptop's app, running on the Mini the whole time.
Measured or publishedFrom the move and the record
Started on the phoneContinued in a browserFinished in the laptop’s app
One conversation, running on the Mini the whole time. Each window was only a view onto it.
All three apps listed the same sessions, because none of them was running Earl. The Mini was.[11]
For whoever runs it: Pick one machine to be the agent. Everything else is a window.
Evidence and objections
Measured The laptop's desktop app lists the Mini as a place to start a session, beside the laptop itself. Sessions started from the phone appear in the laptop's sidebar and the other way around.[11]
Measured Before the move, the phone link depended on the laptop being open, and 1 of the last 235 sessions had it switched on.[5]
f4moderateFour sessions, before and after
Each session starts about 12,500 tokens lighter on the Mini.
MeasuredA session opened from the laptop's app reads about 75,000 tokens before its first reply on the Mini, against about 87,000 on the laptop before.
Measured or publishedFrom the move and the record
On the laptop, beforeone session, less the message I opened it with
about 87,000
On the Mini, afterthree sessions, 74,547 to 74,894
about 75,000
Tokens read before the first reply. About 8,000 of the drop is the in-app browser’s tools, which a Mini session doesn’t load, and about 2,000 is memory recall I switched off the same day.
The size of each session's first request, from Claude Code's own records. Tokens are the chunks a model reads, and fewer means faster and cheaper.[14]
For whoever runs it: Measure the starting load before and after any move. It's the tax every session pays.
Evidence and objections
Measured About 8,000 of the drop is the tools for the browser inside the laptop's app, which a Mini session doesn't load. About 2,000 is automatic memory recall, switched off the same day because a review found it added something new in only 5% of what it pulled in.[5, 14] The last 2,500 or so I can't attribute.
Measured The laptop figure is one session's first request, 96,539 tokens, less the 9,331-token message I opened that session with.[14]
Objection, stated: three sessions on one day is a thin sample. The shake-out week adds more, including sessions started from the phone.
D
Part D: What it reaches, and what it gave up
The reason to do this at all, and the bill for it.
f5strongEvery read-only check, from the Mini
Earl reaches the whole network. A cloud agent reaches one computer, at most.
MeasuredFrom the Mini, every service I checked answered: the media servers, the task board, the storage servers, the local model. The cloud agents can reach one computer you allow, and only while it's awake.
Measured or publishedFrom the move and the record
A cloud agent can reach
The web, through its own browser
The apps you connect: mail, calendar, documents
One computer of yours, if you allow it and it’s awake
Earl on the Mini can reach
The web, and the same kind of connected apps
The media servers and the services that feed them
The storage servers, through a remote login
The task board and the wiki
A local AI model on the Mini itself
What each kind of agent can get to. The home list is what answered a read-only check from the Mini on Oct 22.[12, 7, 3, 9]
For whoever runs it: List what you want the agent to touch before you pick where it lives.
Evidence and objections
Measured OpenAI says a dot's "computer and its contents stay separate unless you choose to connect it," and that you can let it use your laptop. Muse's Mac app "works with your desktop computer files, apps and browser tabs." Grok Bot can run commands on "the computer in front of you."[3, 9, 7] None of the four pages describes reaching anything else on a home network.
The way the cloud would get in: a cloud agent can only reach a home service through a door opened to the internet. Remote Control avoids that: the Mini calls out, and nothing calls in.[8]
f6moderateEvery session's tools, both machines
The app's best features stayed on the laptop.
MeasuredOn the laptop, the desktop app pins a task list under every message and ends each turn with buttons for the likely answer. In 57 sessions on the Mini, neither showed up once. Neither did the browser inside the app.
Measured or publishedFrom the move and the record
At the desk: the laptop's appAway: the Mini, over Remote Control
A task list pinned under each message
Yes
No
Buttons with the likely answer, one click
Yes
No
The browser inside the app
Yes
No. Earl drives its own headless Chrome
Drawings and widgets inside the chat
Yes
No
Computer use and the Chrome extension
Yes
No
Sending a file into the chat
Yes
No. The tool is there and fails, so Earl copies files over
Mods: panels and strips beside the chat
Installed here
Not installed there, so not tested
Reaching it from the phone
Only while the laptop is awake
Always
Published pages, phone alerts, dictation
Yes
Yes
My mail and calendar connections
No
Yes
From the tools each session was given, in Claude Code’s own records: 49 sessions on the laptop, 57 on the Mini.
What each kind of session had, Oct 22 to Nov 3, 2026. "Away" is the Mini reached over Remote Control, from the phone, the web or the laptop's app.[16, 8]
For whoever runs it: Write down what you use every day before you move, and try each one after.
Evidence and objections
Measured Those are the desktop app's own additions, and they come with sessions the app runs itself. A Mini session reached over Remote Control gets Claude Code without them.[16]
Measured Mods, the add-ons that put a panel or a status strip beside the chat, are installed one machine at a time. Mine is on the laptop only, so it never loaded on the Mini. Nothing in Anthropic's documentation says a mod can't run remotely, but installing one is a command that works "only from the local CLI," so it has to happen at the Mini.[8] Whether a mod's panels then show up on the phone, I haven't tested.
Measured Live voice was never on the table. In Claude Code, "dictation is available but voice mode isn't," on either machine.[15]
Measured The Mini had one thing the laptop didn't: my mail and calendar connections from claude.ai showed up in every Mini session and none on the laptop.[16]
E
Part E: What the copy didn't carry
Moving an agent with months of setup turned out to be mostly finding out what that setup depended on.
f7strongEvery item from the checklist
The copy took 46 seconds. Ten things still needed setting up, and two gave no error.
MeasuredThe code moved in one copy, 3.3 GB in 46 seconds. The rest of the move was ten things the copy didn't carry: setup that lived outside the code, found and done one by one. Two of them gave no error to find them by.
Measured or publishedFrom the move and the record
Needed my hands at the screen
Every sign-in. The Mac's keychain won't open over a remote connection.
Didn't travel with the code
179 font, music and sound files (86 MB) kept out of the repository
A 756 MB model file, same reason
Five settings files with keys in them
Wrong, with no error
The bundled Chrome unpacked halfway, with no error
A health check passed the fonts while the font files were missing
Still pointed at the old machine
Preview pages listening only to the Mini itself
The image tool's path inside a newer app
Usage reports still naming the laptop
Too loud
The server's status screen, about 40 MB of log a day
Every step the move needed after the copy, grouped by why.[12]
For whoever runs it: Keep the setup in files you can copy, then check that the copy actually runs.
Evidence and objections
Measured Most of the fixes were small, and Earl made them itself. The sign-ins needed me at the Mini's own keyboard, because a Mac's keychain won't unlock for a remote login.[11, 12]
Measured The setup being plain files in a git repository is what made the move possible in a day. Everything that wasn't in the repository is on this list.
f8moderateOne machine, one power risk
The Mini hasn't restarted since Aug 7. A power cut is the one thing that would stop it.
MeasuredThe Mini has run since Aug 7 without a restart, 87 days as of Nov 3. Its disk is encrypted, so after a power cut it would wait at the unlock screen until someone types the password.
Measured or publishedFrom the move and the record
The power goes outThe Mini boots back upIt waits at the unlock screenSomeone types the password
Up since Aug 7 without a restart. The disk is encrypted, and an encrypted Mac won’t log itself in.
For whoever runs it: Decide ahead of time what wakes you up when the agent goes quiet.
Evidence and objections
Measured It has happened once, in August, before Earl moved in, when nothing urgent lived on the Mini. It waited at the unlock screen for 33 hours.[12]
Measured Turning the encryption off would let the Mini log itself back in, and would leave the disk readable by anyone holding it. I've kept it on.[12]
Still open: a phone alert when the Mini goes quiet, so a stuck unlock screen costs minutes, not a day. Not built yet.
My reading, not measured
The trade I'm least comfortable with is the one in the guardrails row. Earl runs with broad permission on my network. It does things without stopping to ask that every cloud agent on this page would ask about first. I chose that knowingly, because an agent that asks before every command isn't much use from a phone. But it means a mistake, or a booby-trapped web page that steers it, lands on my own machines. The cloud agents' guardrails exist for exactly that, and here, they're up to me.
F
Part F: Two machines, one agent
The move took a day. Living with it took the next twelve, and it changed where I work.
f9moderateEvery message I typed, twelve days
A week in, I split the work by one question: am I leaving?
MeasuredFor the first week, every conversation ran on the Mini. On Oct 29 I moved desk work back to the laptop for the app's features, and kept the Mini for whatever has to run while I'm gone.
Measured or publishedFrom the move and the record
On the Mini On the laptop
Oct 22
3444
Oct 23
410
Oct 24
590
Oct 25
440
Oct 26
430
Oct 27
350
Oct 28
4313
Oct 29desk work goes back to the laptop
2984
Oct 30
092
Oct 31
030
Nov 1
059
Nov 2
04
Messages I typed each day, by the machine the session ran on. The Mini’s last conversation was Oct 29 at 2:59 AM. Since then it runs the jobs on a timer.
Messages I typed each day, by the machine the session ran on. Scripted jobs aren't counted.[17]
For whoever runs it: Decide where a session runs before it starts: will you walk away before it's done?
Evidence and objections
Measured The rule, as I wrote it down at 2:13 AM on Oct 29: start a chat on the laptop "if I know I'm not going to need to step away so I get all the features like mods. Then if I'm going to be stepping away or need something to run while my laptop is closed, I do it on the mini."[17]
Measured What pushed it: four hours earlier, I'd started a job on the laptop that was meant to run overnight on the Mini, and only noticed almost two hours in.[19]
Measured Since that night, every conversation has been on the laptop. The Mini runs the jobs on a timer, like an email brief every morning at 8, and waits for the night I need it. Once since then I left the laptop open on power overnight instead, the habit the Mini was supposed to end.[17]
f10moderateOne split, twelve days
Two copies stay one agent because every session starts by catching up.
MeasuredBoth machines carry the same rules, memory and code. Every session pulls the latest before it starts, and every change is backed up the moment it's saved. It split once, and nobody noticed for 18 hours.
Measured or publishedFrom the move and the record
A session starts, on either machineIt pulls the latest, but only a clean catch-upIt works and saves a changeThe change is backed up to a server at homeThe other machine pulls it at its next start
If a pull is refused, one line says so at the top of the session. It split once in 12 days, on Oct 29, and the Mini’s warning went unseen for 18 hours, because no session opened there to show it.
The loop that keeps the two copies in step, as it stands on Nov 3.[18]
For whoever runs it: Build the sync before you need it, and let only one machine save at a time.
Evidence and objections
Measured The split, Oct 29: the Mini saved three notes at 3 AM after a session wrapped up. The laptop picked up the same work in the morning and saved first. From then on the Mini's catch-ups were refused. The warning only shows when a session opens, and no conversation opened on the Mini until the evening, while 47 laptop changes piled up. One merge that evening fixed it, and nothing was lost.[18]
Measured The rules it left behind: the end-of-session jobs never run on both machines at once, and a note handing work to the next session is claimed by one session before it starts. Two sessions had picked up the same note three minutes apart on Nov 1.[18]
What it doesn't fix: when the backup server at home is down, the warning line is all you get. That happened twice.
f11moderateEvery snag in the session notes
The second machine needed its own upkeep.
MeasuredMost of the friction came from the Mini needing care like any computer: memory, housekeeping, and files that ended up on the wrong side.
Measured or publishedFrom the move and the record
Ran out of memory
14 sessions open on the Mini, when I meant to have two, and a memory helper grown to 3.5 GB.
The helper restarts itself when it grows, and wrapping up a session now closes it.
Hit the session cap
At about 30 open sessions, no new one would start until I archived some.
Still by hand, in the browser. Earl's own archive tool waits on a prompt that never appears.
Fell behind
The Mini's copy of the setup fell 27 changes behind, held up by files the Claude app rewrites itself.
“Update the Mini” is still a request I make every few days.
Files on the wrong machine
Eight requests in one week to put a file on the laptop, from sessions running on the Mini.
A direct copy over the home network, straight into the laptop's Downloads.
Started on the wrong machine
I opened a job on the laptop that was meant to run overnight on the Mini, and found out almost two hours in.
Four hours later I set the rule: the laptop at the desk, the Mini when I leave.
Each snag from the twelve days after the move, and where it stands.[19]
For whoever runs it: Budget time to look after the second machine, like any other computer.
Evidence and objections
Measured None of these lost work. Each cost an evening or a redo, and each is now a rule or a fix in the setup, except the session cap.[19, 18]
Objection, stated: twelve days is a short run, and the Mini has carried no conversations for the last five of them. The night shift hasn't been tested much since the split.
My reading, not measured
What I'd tell someone setting this up: give the always-on machine the night shift and the phone line, and keep desk work on the laptop, because the best of the app lives where the app runs. Hand off to the home machine when you leave. The sync is what makes the handoff safe.
A fit, and not a fit
A good fit if
You already run things at home an agent should reach: servers, file shares, local models, home automation.
You have a machine that's always on anyway.
You're comfortable being your own IT department, including at 2 AM.
Your agent already has a setup worth keeping, in files.
You work at a desk most of the day and still want the agent when you're not.
Not a fit if
Your work lives in email, calendars and web apps. A cloud agent reaches those already.
You want the guardrails done for you.
You'd want every app feature wherever you are. Today the best ones stay with the machine the app runs on.
Nobody will be home to type a password after a power cut.
Four decisions you'll make
How much the agent may do without asking.
More permission makes it useful from a phone. It also makes a mistake land on your own machines. Decide this on purpose, not by default.
Desk work where the app runs, anything that outlives your attention on the always-on machine. Decide it at the start of each session, not halfway through.
A catch-up that refuses anything but a clean one, run before each session, plus a backup on every save. The one split came the night two machines saved at once.
That this is better than a cloud agent. It answers a different question.
That I tested the four cloud agents. I compared what their makers say. I haven't used any of them, and Duet's waitlist hasn't reached me.
That the model runs at home. It doesn't. The model is in Anthropic's cloud, exactly as before.
That it's safe. Its guardrails are the ones I built, and they haven't been tested by anyone but me.
That it holds up for months. This is the day of the move and the twelve after.
That the feature list is complete. It covers what my sessions were given, and mods weren't tried on the Mini.
Written down first
Before anything moved, there was a written plan: what had to move and how, what I'd have to do at the Mini's screen, and a checklist to work top to bottom.[12] Here's how its assumptions held.
“Only one machine works at a time until cutover.”
Kept
Both copies back themselves up to the same place, so two writers would have split the history. They never overlapped.
“Everything Earl needs is in the repository.”
Mostly
The code and the setup were. Fonts, music, a model file and five settings files weren't.
Every idea tried during the move, and what became of it.
Killed
Automatic memory recall. A review found it added something new in about 5% of what it pulled into each session. Switched off. Earl searches its memory when it needs to.[5]
Checking it's alive by messaging it. Thirty checks cost 33 model calls. The Mac's own process list is free.
The trial copy. Nothing unique in it, so it was deleted once the real copy ran.
Every conversation on the Mini. A week in, desk work went back to the laptop for the app's features (f9).
A timer on the laptop for a usage report. It had stopped in June without anyone noticing, and the same report already runs at the end of every session.
Changed
Preview pages. They listened only to the machine they ran on, so the laptop couldn't open them, including the review page the agent opens beside the chat. Now they listen on the home network.
The bundled Chrome. Unpacked by hand after its own installer left it half-done.
The server's log. Its status screen redrew itself twice a second into the log, about 40 MB a day. Sent nowhere now.
Two tools' web addresses. They pointed at the public address; now the home one.
The sync. A catch-up at every start, a stop at wrap-up if it's refused, and a claim on each handoff note (f10).
Held up
The lid test. 30 of 30 (f1).
The crash test. Back in 2 seconds (f2).
One agent, three windows (f3).
The backups. Every change on the Mini reached the backup server, signed, and the Oct 29 split lost nothing.
The day
Oct 21Anthropic announces Claude Duet
11:58 PMThe review starts: what Earl is, and where it should live
1:30 AMA trial copy goes onto the Mac mini
2:45 AMThe trial passes: lid closed, crash, three windows
MorningThe real move, mostly done by Earl over a remote connection
12:15 PMThe always-on server switches to the real copy
12:30 PMThe shakedown checklist, section by section
4:05 PMCutover: the laptop stops being Earl's home
The move, from the commit history and the checklist. Times are Eastern.[12]
Limits
One move, one Mac, one network. The day of the move, and twelve days after it.
The cloud agents weren't tested. Every claim about them is what their makers say, and three cells rely on press coverage, marked "reported."
Small samples. One lid test, one crash, three sessions for the starting load. The feature list rests on 49 laptop sessions and 57 on the Mini.
Phone and laptop look the same in the Mini's records. Sessions started from either can't be told apart, so the feature list treats them as one.
No restart or power-cut test yet. The Mini has been up since Aug 7; its one power-cut wait, in August, was before the move.
Mods weren't tried on the Mini. None was installed there.
How it was made
Personal time, personal hardware, public pages. The trial ran from 1:30 to 2:45 AM ET on Oct 22, 2026, on a separate copy that couldn't write back to anything. The real move ran the same day, and the cutover was at 4:05 PM.[11, 12] The Mini runs Claude Code's Remote Control in server mode, kept alive by the Mac's service manager.[8] The four products' pages were read on Oct 22. OpenAI's and Meta's were read with a headless browser, because they turn away plain fetches. Part F was read on Nov 3 from both machines' session records and commit history: the tools each session was given, the messages I typed each day, and the notes from each snag.
AI disclosure
The prose here was written by an AI, Earl (Claude Opus 5.5), under Rick's editorial direction. It's also Earl's own move: Claude wrote the plan, did most of the migration over a remote connection, ran the tests and drew the figures. Claude also went through both machines' session records for Part F. Rick made every decision, did the sign-ins at the Mini, tested the laptop's app against the new setup, set the laptop-at-the-desk, Mini-when-away rule from his own use, and owns every editorial call.
So you're trusting my records of the move and the four companies' own descriptions of their products. Each is named under Sources.
Versions
Rev 1 First publication, Nov 3, 2026. Drafted Oct 22, the day of the move; Part F adds the twelve days after it.
SourceDigital Applied. "Grok Bot: xAI's AI Teammates Get Their Own Computer," read Oct 22, 2026. Secondhand: launched in beta Aug 11, 2026; access through SuperGrok Heavy or Cursor's top tiers (Cursor Ultra $200 a month); xAI's docs say all of an account's Bots share one cloud computer, against the launch page's "Bots have their own computer." www.digitalapplied.com/blog/grok-bot-ai-teammates-launch-cloud-computer-2026grok-launch
SourceLet's Data Science. "Meta Launches Muse Personal AI Agent," read Oct 22, 2026. Secondhand: launch date Sep 8, 2026, a hosted cloud computer on Meta's infrastructure, and a free tier with $20 and $100 monthly plans in the US. Not confirmed on Meta's own page. letsdatascience.com/news/meta-launches-muse-personal-ai-agent-403f80e4muse-launch
SourceOpenAI. "Introducing dots," Sep 29, 2026, read Oct 22, 2026. Quoted: dots "have their own cloud computer… and can work towards your goals 24/7"; "Each dot works on its own cloud computer, while your computer and its contents stay separate unless you choose to connect it"; "you can also give your dot permission to connect and use your laptop"; "hop on a voice call". Background work uses read-only tools; auto-review and Custom Rules govern actions. First dot included in Pro and Business Premium. openai.com/index/introducing-dotsdots
SourceAnthropic. Claude Duet product page and launch article (claude.com/resources/articles/claude-duet), Oct 21, 2026, read Oct 21–22, 2026: a separate app where Claude works on its own cloud computer, checks connected accounts "every few hours", takes tasks by text or recorded voice message, and asks before sending, paying, signing or deleting. Research preview, waitlist, US only, Pro and Max plans, sharing the plan's usage limits. claude.com/product/claude-duetduet
DerivedMy architecture review of Earl, Oct 21–22, 2026 (unpublished, my own records): five options scored on seven criteria, and the session counts behind "before": 1 of the last 235 sessions had Remote Control, the phone link, switched on, and 27% of sessions opened by re-explaining context (35% in October).review
SourceApple Support. "Mac mini power consumption and thermal output (BTU) information," read Oct 22, 2026: Mac mini (2024) with M4, 16 GB, 4 W idle and 65 W maximum, measured at the wall. support.apple.com/en-us/103253apple-power
SourcexAI. Grok Bot documentation, "Approvals, security and privacy," read Oct 22, 2026. Quoted: "The shared Grok Bot computer runs in the cloud"; "Access to the Mac or Windows computer in front of you is a separate capability" (setting "Execution on Local Computer", default "Ask every time"); the page advises "Never allow unless a Bot has a specific reason to work on your local files." docs.x.ai/grok-bot/approvals-security-and-privacygrok-docs
SourceAnthropic. Claude Code documentation, "Remote Control," read Oct 22 and again Nov 3, 2026: a Claude Code session running on your own machine, reached from the Claude app on a phone, the web or the desktop app. The machine connects out to Anthropic; nothing on it is opened to the internet. Quoted (Nov 3): commands "such as `/plugin` or `/resume`, work only from the local CLI." code.claude.com/docs/en/remote-controlremote-control
SourceMeta. Muse product page, read Oct 22, 2026. Quoted: "A secure computer with a browser shared by you and your agent"; "Review and approve actions from Muse before they happen, like sending emails and making purchases"; "Muse works with your desktop computer files, apps and browser tabs" (the Mac app). The page gives no launch date or price. ai.meta.com/musemuse
SourceAnthropic Help Center. Claude Duet help article, read Oct 22, 2026. Quoted: it "can't see all your projects or files on your computer, and it can read but not change your work in Claude." support.claude.com/en/articles/17631495duet-help
DerivedMy trial log, Oct 22, 2026, 1:30–2:45 AM ET (my own records): the lid-closed test (30 one-a-minute heartbeats, no gap through 11 minutes 47 seconds of laptop sleep), the forced crash and restart, the one session moved across phone, browser and desktop app, and each window's starting load.trial
DerivedMy migration plan and shakedown checklist, Oct 22, 2026 (my own records), each item ticked with what was found, and the commit history that timestamps every step from the first copy to the 4:05 PM cutover.shakedown
DerivedThe Mac mini's own uptime and boot time (my own records): up 76 days when read on Oct 22, 2026, 8:25 PM ET, and 87 days on Nov 3, 1:15 AM ET. Last booted Aug 7, 2026, 4:36 PM ET.uptime
DerivedClaude Code's own session records on the laptop and the Mini (my own records): the token count of each session's first request, read from the stored usage figures. Before: 96,539 on the laptop, of which 9,331 was the opening message, 2,075 automatic memory recall and about 7,917 the in-app browser's tools. After: 74,547, 74,619 and 74,894 on the Mini, Oct 22, 2026.sessions
SourceAnthropic Help Center. Voice mode article, read Oct 22, 2026. Quoted: "In Claude Cowork and Claude Code, dictation is available but voice mode isn't." support.claude.com/en/articles/11101966voice
DerivedClaude Code's own session records on both machines, Oct 22 to Nov 3, 2026 (my own records): the tools each session was given, read from its first request. 49 desktop-app sessions on the laptop, 57 Remote Control sessions on the Mini. No Mini session ever called the app's task-list or answer-button tools or its built-in browser; every Mini attempt to send a file into the chat failed.features
DerivedMy typed messages per day per machine, Oct 22 to Nov 2, 2026, Eastern days (my own records, from both machines' session files): scripted prompts and copies from resumed sessions dropped. The last conversation on the Mini ended Oct 29, 2026 at 2:59 AM ET.days
DerivedMy sync records, Oct 22 to Nov 3, 2026 (my own records): the start-up hook that pulls both repositories (added Oct 29, 2:18 AM ET), the backup on every commit, the Oct 29 split (three Mini changes unbacked for about 18 hours under 47 laptop changes) and its one-merge repair at 9:11 PM, and the handoff-note claim added Nov 1.sync
DerivedMy session notes from the twelve days after the move (my own records): the Mini out of memory (Oct 29, 1:44 AM ET), the session cap (Oct 26), the catch-ups (Oct 30 to Nov 1), the file copies to the laptop (Oct 23 to 28), and the job started on the wrong machine (Oct 28, 9:56 PM).upkeep