One Mac or Two Machines? Choosing the Right Home Layout for Your Local AI Model

James Ballard · September 26, 2026

This is post #8 in our series on running your assistant's everyday replies on an AI model on your own Mac. So far we've covered:

This post covers a decision that comes up before any of that: where do the pieces physically live? Launchpad supports two home layouts for a local model. Most people should pick the first one, but the second is a good fit for a specific kind of household.

A quick reminder first. "Bring your own AI model" is an advanced-users-only feature. It works with an assistant that runs on a computer you own. It's a dashboard setting, not a separate plan.

The two layouts, in plain English

Two parts work together when you use a local model:

Your assistant. This part receives your Telegram, WhatsApp or email messages, keeps track of the conversation, and works with your approved Google capabilities. Dex, your built-in AI sysadmin, looks after it and keeps things healthy.
The AI model. This is the "brain" that writes the replies. It runs in the free oMLX app on an Apple-silicon Mac.

Launchpad supports exactly two ways to arrange them:

Layout A: One Mac. The assistant and the model both run on the same Mac. This is the simplest option, and it's the one we recommend.
Layout B: Two machines, one home network. The assistant runs on another computer in your home, for example a Raspberry Pi. The model runs on your Mac. The two talk over your home wi-fi.

There's no Layout C. You can't reach your Mac's model from a cloud-hosted assistant or over the internet. That option is switched off in the product.

Why we recommend one Mac for most people

Layout A has fewer moving parts. Every extra machine is one more thing that can be unplugged, updated or asleep at the wrong moment.

With one Mac:

There's one device to keep awake and online. Our Mac setup help article suggests a Mac mini as an always-on option. It also warns that closing a MacBook's lid usually puts it to sleep, and a sleeping Mac can't answer anyone.
Replies are written on the Mac itself. When the local model is answering, your messages aren't sent to a cloud AI model to produce the reply.
There's one thing to check after a restart. You start oMLX by hand, and nothing restarts it for you after a reboot. We learned this firsthand: our own test Mac sat dead for six days because oMLX wasn't reopened after a restart. With one Mac, you only have one "did I reopen everything?" checklist.

To run the assistant itself on a Mac, you need macOS 14 (Sonoma) or newer. For memory, a large model takes about 15GB, which is comfortable on a 32GB Mac. A 16GB Mac only works with a noticeably smaller model. The hardware guide goes into detail.

When two machines make sense

Layout B is worth considering in these situations:

You already run your assistant on a Raspberry Pi or another spare computer. You can add a local model without rebuilding everything on the Mac.
Your Mac is your everyday work computer. You might prefer it to only lend its processing power for replies, while the assistant lives on a small machine that stays out of your way.
You want routine upkeep to carry on while the Mac is off. With the assistant on a separate machine, the Mac being closed or asleep doesn't take the assistant's computer down with it.

Be clear about what Layout B does not give you: replies still stop when the Mac is unavailable. If the Mac sleeps, loses its network connection or the model stops, your assistant stops answering. It doesn't switch back to Claude on its own, in either layout. A separate assistant computer doesn't change that.

What you need for the two-machine layout

A computer for the assistant. Launchpad's supported self-hosted options include:
a Raspberry Pi 4 or 5 (4GB or more)
a spare Linux machine
a Windows 10/11 PC
a Mac

Setup is one copy-paste command. On Windows and Mac there's a one-time step of about 10 minutes first. See our BYOD announcement for the overview, or the run-on-your-own-computer page for the current requirements.

An Apple-silicon Mac with enough memory, with oMLX installed and a model downloaded and running. The dashboard reads the model list from oMLX, so you don't have to type any model names.
oMLX's API key. Think of it as the password your assistant uses to talk to oMLX. You'll find it inside oMLX under Security.
Your Anthropic key, still connected. Dex keeps using Claude for its maintenance work at a small cost, so the key is required in both layouts.
Both machines on the same home network.

One important note: where your assistant lives is decided when you create it. If you have a cloud-hosted assistant today, it can't be moved onto your hardware. To use a local model, you'd create a new assistant on a computer you own.

What changes when something goes wrong

The failure points differ between the two layouts, so it helps to know them in advance.

| What happens | One Mac | Two machines | |---|---|---| | Mac sleeps or shuts down | Assistant and model both stop | Assistant stays up, but can't get replies | | Mac restarts | You reopen oMLX by hand | You reopen oMLX by hand | | Wi-fi hiccup | Your messages may not reach the assistant | Can also cut the link between the two machines | | Assistant computer loses power | Same as Mac going down | Assistant stops, even if the Mac is fine |

How you find out: in our testing, you'll usually notice first. A message fails after about 3 minutes. Our monitoring detects the problem within about 12 minutes and sends one email.

There's a gap in the one-Mac layout. If the Mac itself goes to sleep, you get no email at all. The failed message is your only alert.

The one-click escape hatch: in either layout, you can switch back to Claude from the dashboard in one click.

Switching back to Claude starts a fresh conversation.
Switching to the local model keeps your conversation.

The recovery guide walks through the full process of getting a stalled model back.

What changes for privacy

In Layout A, replies written by the local model are produced entirely on your Mac.

In Layout B, your conversation travels across your home network between the assistant's computer and the Mac. Both machines are yours and in your home, but more of your household network is involved. If you share your wi-fi widely, keep that in mind.

In both layouts:

Your messages still arrive through Telegram, WhatsApp or email, which are outside services.
Dex still sends its own maintenance work to Claude.

Our privacy map breaks down what goes where.

What changes for cost

Here's what stays the same:

Your subscription. Using a local model doesn't change your plan.
Your Anthropic key. It's still connected, and Dex still uses a little Claude credit.

Here's what changes:

No per-message model charges. Your conversations are free apart from electricity.
A second always-on device. In Layout B, you're powering two machines instead of one. A Raspberry Pi uses very little electricity, but it isn't zero.

Running on your own hardware also means there's no cloud server bill. It isn't a cheaper Launchpad plan, though. It's just a different place to run your assistant. For the full picture, see what a private assistant really costs per month.

What isn't supported, in either layout

So there are no surprises:

Only oMLX is built and tested as the model app. Other model apps aren't supported.
Only Apple-silicon Macs can run the model. Windows PCs, Linux machines and Raspberry Pis can host the assistant but not the local model. Intel Macs can't run the local model either. Check the run-on-your-own-computer page for which computers can host the assistant.
There's no internet or remote access to your model.
We support the connection, not your Mac's hardware or the model you choose. Some models, in our testing, are confidently wrong or struggle with multi-step jobs. Pick carefully, and use Claude for the tasks that matter.

On speed, our testing on an M-series Mac averaged about 10 seconds per reply, with a range of 5 to 46 seconds. The first reply after the model loads can be much slower. Your results will depend on your machine and model. We haven't published separate figures for the two-machine layout.

A quick decision checklist

Choose one Mac if:

You're setting things up fresh.
You have a Mac that can stay awake, such as a Mac mini.
You want the fewest things to check after a restart.

Choose two machines if:

Your assistant already runs on a Pi or another computer you own.
Your Mac is your daily work machine and you'd prefer to keep the assistant separate.
You want the assistant's computer to stay up while the Mac is off, and you accept that replies still stop when the Mac does.

Either way, the core idea stays the same: you own the machines, the accounts and the conversations, and Dex handles the upkeep of your assistant so you don't have to open a terminal. When you're ready, the bring-your-own-AI page and the setup guide will take you through it.