What Actually Stays on Your Mac When You Run a Local AI Model? A Plain-English Privacy Map

If you point your assistant at a model on your own Mac, you're probably not chasing a gadget. You're after control: your day-to-day conversations handled by hardware you own, not sent off to a cloud AI service for every reply.

That's a real benefit. It's also easy to overestimate. "Local" describes one important part of your setup, not every part of it. This seventh post in our Mac local AI series maps out, piece by piece, what stays on your Mac and what still travels. That way you can make decisions based on what actually happens.

If you're new to the series, start with the overview of bring-your-own-model on Mac, then the hardware guide and the step-by-step setup guide.

The Short Version

| Part of your setup | Where it happens when you use a local model | |---|---| | Your assistant working out its replies | On your Apple-silicon Mac, in oMLX | | Traffic between assistant and model (two-machine setup) | Across your home network | | The message you send and the reply you get | Through the app you used (Telegram, WhatsApp, or Gmail) | | Dex's maintenance and repair work | Still uses Claude, through your own Anthropic key | | Google Calendar, Contacts, Drive (if connected) | At Google, as always | | Off-server backups (if set up) | Your Google Drive |

The rest of this post explains each row.

What Stays at Home

Your assistant does its thinking on your Mac

This is the heart of the feature. When "Bring your own AI model" is switched on in your dashboard, your assistant sends its day-to-day replies to a model running in oMLX on your own Apple-silicon Mac. It doesn't send them to Claude. The model reads your message, works out a response, and hands it back, all on hardware you own. There's no per-message AI charge for those replies.

Picking the model is your job, and you don't have to type a model name. The assistant reads the list of available models from oMLX. That choice matters for privacy too, because you decide which software handles your conversations.

In a two-machine setup, the traffic stays on your home network

Two setups are supported today:

• One Mac. Your assistant and the model both run on the same Mac. This is the setup we recommend because it's the simplest.
• Two machines in the same home. For example, your assistant runs on a Raspberry Pi or a spare computer, and the model runs on a bigger Mac on the same wi-fi.

In the two-machine version, your assistant and your Mac talk to each other across your home network. Reaching your model over the internet is not something the product offers. That option is switched off. So if you've seen guides elsewhere about opening up your home model to the outside world, they don't apply here, and we don't recommend following them.

No quiet detours to Claude

This one may surprise you. If your Mac sleeps, loses its network connection, or oMLX stops, your assistant stops answering. It does not switch over to Claude on its own to cover the gap.

That can be frustrating, and restarts are the most common cause. oMLX doesn't reopen by itself after your Mac reboots, so you'll need to start it again. Our recovery guide walks through getting things back. From a control standpoint, though, this is the behavior you want. Your assistant won't hand your conversation to Claude on its own just because your Mac went quiet. Switching back is your decision.

What Still Leaves Your Mac

Your messages travel through the app you use

A local model changes where replies are worked out. It doesn't change how messages reach your assistant. If you message your assistant on Telegram or WhatsApp, or email it through Gmail, the message and the reply still pass through that service, the same way any message you send in those apps does.

That's worth being clear-eyed about. You may find it an acceptable trade if you already use these apps every day and the part you want to keep away from a cloud AI service is the AI processing itself. If a specific conversation is especially sensitive, think about the whole path, not just the model.

It's also worth remembering that your assistant only answers contacts you've approved. That boundary stays in place whether you're on Claude or a local model.

Dex still uses Claude

Your Anthropic key stays connected even when your conversations run locally. Dex, the built-in AI sysadmin that monitors, patches, backs up and repairs your setup, keeps using Claude for its own work. That costs far less than running your conversations through Claude, but it isn't zero, and it's billed to your own Anthropic account. So a local-model setup is not "completely offline" or "completely free." Electricity for your Mac is a cost too.

Google connections and backups live at Google

If you've connected Google Calendar, Contacts or Drive, each approved separately, that information lives where it always has: in your Google account. A local model doesn't pull it out of Google.

Off-server backups, which need a writable Drive connection, are stored in your Google Drive. That's intentional. A backup is only useful if it survives something going wrong with the machine. Just know that "off-server" means "off your Mac."

What LaunchMy.ai Can and Can't Do

LaunchMy.ai is a deployment launcher, not an operator. Your assistant runs on hardware you own. Your AI usage is billed to your own Anthropic account, and we don't hold the keys to your machine. Your dashboard shows live status and logs for your assistant, and we can tell whether your assistant can reach your Mac. We don't support your Mac itself or the model you've chosen. You pick them, so you own them and the answers they produce.

We'd rather give you that plain description than a sweeping promise that "nobody can see anything." Ownership is the point: your assistant runs on your hardware, and your AI usage runs through your own accounts.

Switching Between Local and Claude

Switching is one click in the dashboard, in either direction. It's a setting, not a plan change, so your subscription stays the same.

The two directions behave differently:

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

From then on, your assistant's replies are handled by Claude through your Anthropic account again. If you expect to switch back and forth (our post on when to use a local model versus Claude can help you decide), keep that in mind when you decide which topics to raise after switching.

What "Local" Doesn't Mean

To avoid surprises, here's what this setup is not:

• Not available for a cloud-hosted assistant. Your assistant has to run on your own computer at home (a self-hosted setup), a choice made when the assistant is created. A cloud assistant can't be moved onto a Mac later. You'd set up a new assistant on your own hardware. Our BYOD announcement covers that option.
• Not for any Mac. It needs Apple silicon with enough memory. A large model needs about 15GB on its own. A 32GB Mac handles that comfortably, and a 16GB Mac only works with a noticeably smaller model.
• Not unmonitored. If something goes wrong, you'll usually notice first, when a message fails after about 3 minutes. Our system notices within about 12 minutes and sends one email. On a one-Mac setup, a sleeping Mac means no email at all, because the part that would send it is asleep too.
• Not fast in the way cloud AI is. In our testing, replies typically took about 10 seconds, ranging from 5 to 46 seconds on an M-series Mac. Your results depend on your machine and model.

A Privacy Checklist for Local-Model Owners

1. Know which setup you're on: one Mac, or two machines on your home network.
2. Choose your chat app deliberately, since messages travel through it.
3. Connect only the Google capabilities you actually need. Each one is approved separately.
4. Decide whether off-server backups to Drive fit your comfort level.
5. Remember Dex still uses your Anthropic key, and keep an eye on its usage in your reports.
6. Reopen oMLX after every restart. Otherwise your assistant simply stops answering. It won't switch to Claude on its own.
7. Pick a capable model for multi-step jobs. Smaller models can struggle with work involving files or pictures, and a less capable model once reported finishing a step it hadn't.

For more on routines that make this manageable, see living with a local AI model day to day.

The Bottom Line

Running your assistant on a model on your own Mac gives you real control over the most sensitive part of the process: the AI handling your everyday conversations. It doesn't make every piece of your setup local, and it isn't meant to. Your chat app, Dex's maintenance work, and any Google connections still travel beyond your Mac. Knowing where each piece lives is what lets you use this feature with confidence.

The feature is labelled "Advanced users only" for good reason. If you're comfortable with that, you can read the full details on the Bring your own AI page. It's a setting, not an add-on. LaunchMy.ai plans start at $5.99/month for your first 6 months, then $9.99/month. Your Anthropic usage and electricity are paid separately, directly by you.