# On-device AI apps for Mac

On-device AI on a Mac means the model runs on your own silicon rather than in a data center, which buys you privacy and offline use at the cost of the breadth a frontier model gives you.

- URL: https://swiftfox.studio/blog/on-device-ai-apps-for-mac
- Covers: https://swiftfox.studio/apps/plexus, https://swiftfox.studio/apps/nomore, https://swiftfox.studio/apps/sweep
- Published: 2026-07-30
- Updated: 2026-09-05
- Publisher: Swiftfox Studio (https://swiftfox.studio)

## In short

- "On-device" bundles three separate claims: where the model runs, whether the app makes any network calls at all, and what happens to your content afterwards.
- Apple's Foundation Models framework on macOS 26 gives any app a local model with no API key and no per-token bill, which is what finally made AI viable in one-time-purchase utilities.
- The local model has a small context window, so on-device features are narrow and well shaped by design; a data center model still wins on long documents and hard reasoning.
- Plenty of local decisions need no model at all, and a rule you can read beats inference wherever the job is deterministic.
- You can verify a claim yourself: look for an account, read the privacy label, pull the network, and watch the outbound connections.

## What "on-device" is actually a claim about

"Runs on-device" is three separate promises wearing one label, and an app can
keep any one of them while breaking the others.

The first is **where the model runs**. A local model means the text is processed
by your own silicon rather than shipped to a data center.

The second is **whether the app talks to anything at all**. These two come apart
more often than you would think: an app can run a local model and still upload
your document for search, sync or analytics.

The third is **what happens to your content afterwards**. Retention and training
policies are about a copy that already left. If nothing left, there is nothing to
have a policy about.

Only the third is settled by reading a privacy page. The first two you can check
yourself, and the second-to-last section here is how.

## What Apple Intelligence changed on the Mac

Until recently, putting a language model inside a small Mac app meant renting
one: an API key, a per-token bill, and every user's text crossing the network.
That economics ruled out the whole category of one-time-purchase utilities, which
is why almost none of them had any AI in them and the ones that did were
subscriptions.

macOS 26 ships Apple's Foundation Models framework, which hands any app a
language model that is already on the machine. No key, no bill, no request. It
needs Apple Silicon and Apple Intelligence switched on in System Settings.

The catch is size. The on-device model has a small context window, and apps built
on it have to budget against it rather than pretend it is not there.
[Plexus](https://swiftfox.studio/apps/plexus), for instance, works to roughly a four-thousand-token
window, caps a single call at forty tasks and batches anything larger. That
constraint is exactly why on-device features tend to be narrow and well shaped
rather than an open chat box: a small model doing one specific job is reliable,
and the same model asked anything at all is not.

## Where the cloud still wins

Worth saying plainly, because a comparison that only flatters the local option is
no use to anyone actually deciding.

A frontier model in a data center beats an on-device one at nearly everything
measured in isolation: long documents, current information about the world,
multi-step reasoning, code, images. If you need a fifty-page contract summarized
or a hard problem thought through, ChatGPT, Claude, Notion AI or Raycast AI will
do it better, and no local model on a laptop is close.

What you pay for that is the upload. So the question is not which is smarter. It
is whether this particular text is something you are willing to send. For a
marketing draft, usually yes. For a client's brief, an unreleased roadmap or
anything under an NDA, often no, and at that point the real comparison is not
local model against cloud model. It is local model against doing it by hand.

## An on-device model: Plexus

[Plexus](https://swiftfox.studio/apps/plexus) is a task app built entirely on that framework. You type
one-line todos, the model works out what depends on what, and the canvas rings
the tasks nothing is blocking. Paste a paragraph and it extracts the tasks, names
the flow and wires the ordering in one pass. Ready then collects everything
nothing is blocking across all of them, and Plan my day fits today's into the
hours you actually have.

![The Plexus main window with the sidebar of flows, six colour-tagged task cards wired into a dependency graph, a reference file pinned beside one of them, and the inspector open on Fix the crash on export](https://swiftfox.studio/plexus/listing/shot-1.jpg "The wiring is done by a model on the Mac, and the app makes no network calls of its own.")

Beyond license checks, update requests and anonymous crash and usage counts it
makes no network calls, asks for no permission but notifications, and keeps its
store inside the app container. There is a
feature-by-feature walkthrough in
[turning notes into tasks on a Mac](https://swiftfox.studio/handbook/turn-notes-into-tasks-on-mac). It is $10.99
one-time and needs macOS 26, which is the framework's price of entry rather than
a choice; the binary is universal, and on a Mac where Apple Intelligence is not
available the AI half goes quiet and the rest keeps working.

## Local decisions without a model at all

Most of what people want from "local AI" is not a model. It is a decision made on
their own machine, and for a lot of jobs a rule does that better than inference
does, because you can read a rule and predict it.

[NoMore](https://swiftfox.studio/apps/nomore) blocks distracting sites and apps by checking the
frontmost app and the current browser tab against a list you wrote yourself. The
check happens on your Mac and no URL is sent anywhere. There is no model involved
and there does not need to be, because you already know what you are trying not
to open.

![The NoMore friction panel reading 8 days without Reddit, with Step back as the primary button and Open anyway as a small link below it](https://swiftfox.studio/nomore/listing/shot-2.jpg "The decision is a list you wrote, evaluated locally. No classifier, no upload.")

[Sweep](https://swiftfox.studio/apps/sweep) files downloads by plain conditions, meaning type, name,
size and age, and shows you every planned move before it touches anything.

![The Sweep preview dialog reading Preview - 8 moves, 1 skipped, listing each file with its source and destination and an Overwrite warning on one row](https://swiftfox.studio/sweep/listing/shot-2.jpg "A dry run you can read, which is the thing a model cannot give you.")

That preview is the whole argument. A model sorting your files would be more
flexible and strictly less trustworthy, because you could not check its reasoning
before it moved two hundred things. Where the task is deterministic, determinism
is a feature rather than a limitation.

## How to check the claim yourself

Four checks, in ascending order of effort.

**Look for an account.** An app with no sign-up has nowhere to put your data even
if it wanted to. It is the cheapest signal and the most reliable one.

**Read the privacy label** on the App Store listing, or the privacy page for a
direct download. "Data not collected" is a claim the developer has put in
writing.

**Pull the network.** Turn off Wi-Fi and use the feature. If it still works, the
processing is local. If it hangs, it was not.

**Watch the connections.** Little Snitch, or the free LuLu, will show you every
outbound connection an app opens. License and update traffic is normal and
expected; your documents going somewhere is not.

## Side by side

All three of these are native macOS apps with a one-time price, no account and no
document leaving the Mac. What differs is what makes the decision.

|  | Plexus | NoMore | Sweep | Cloud assistants |
| --- | --- | --- | --- | --- |
| What decides | An on-device model | Your block list | Your rules | A data center model |
| Where it runs | Your Mac | Your Mac | Your Mac | Remote |
| Needs an account | No | No | No | Yes |
| Works offline | Yes | Yes | Yes | No |
| macOS | 26 or later | 14 or later | 14 or later | Varies |
| Price | $10.99 one-time | $8.99 one-time | $9.99 one-time | Usually monthly |

_Prices and features checked on 5 September 2026._

## Frequently asked questions

### What does on-device AI actually mean?

That the model processes your text on your own Mac instead of sending it to a server. It is worth separating from two neighboring claims, though: an app can run a local model and still upload your document for sync or analytics, and retention policies only matter for a copy that already left. Where the model runs, whether the app makes network calls, and what is kept afterwards are three different questions.

### Which Macs can run Apple's on-device model?

Apple Silicon running macOS 26 or later, with Apple Intelligence switched on in System Settings. Intel Macs cannot run it at all. Well-built apps degrade rather than break: Plexus, for example, keeps every manual affordance working and only takes its three AI entry points offline.

### Do these apps work with no internet?

Yes, and that is the practical test. A local model is already on the machine, so the feature works on a plane exactly as it does at a desk. Turning off Wi-Fi and using the feature is the fastest way to find out whether an app's processing is really local.

### Is an on-device model worse than ChatGPT?

For breadth, clearly yes. A frontier model handles long documents, current world knowledge and multi-step reasoning far better than anything running on a laptop. For a narrow, well-defined job that a small model has been prompted carefully for, the gap mostly disappears, and you no longer have to decide whether the text was safe to upload.

### Does an app need AI to keep my data local?

No, and most do not. Blocking a site, filing a download or timing a task are deterministic jobs that a rule handles better than inference, because you can read a rule and predict what it will do. NoMore and Sweep both make their decisions on the machine with no model involved at all.

### How do I verify an app is not uploading anything?

Four checks, cheapest first. An app with no sign-up has nowhere to put your data. The App Store privacy label is a claim the developer has made in writing. Turning off Wi-Fi tells you whether the feature is really local. And Little Snitch or the free LuLu will show you every outbound connection, where license and update traffic is expected but your documents are not.