XDA made the point this week that every modern Mac already has a free, local, on-device language model, and most users have not noticed. Apple Intelligence’s Foundation Models framework ships on every Apple Silicon Mac running Sequoia or newer. Third-party apps can call the local model with a single API, no OpenAI key, no Anthropic account, no data leaving the machine. Seven macOS apps that use the on-device Foundation Models well are worth installing today.
We tested each on an M2 MacBook Air (16GB) and an M3 Pro MacBook Pro. The Foundation Models on-device call runs in about a second for short prompts and stays under a memory footprint that lets us keep other apps open. Cloud fallback to Private Compute or an external provider is available when a query is too heavy.
What to look for in an on-device AI app for macOS
- Foundation Models direct call. Apps that route through the local model, not just OpenAI.
- Global hotkey. The overlay pattern is what makes the local model actually useful.
- Prompt library. A saved list of prompts is what turns a chat window into a workflow.
- Context awareness. Reads the current selection, the current window, or the whole screen.
- Model routing. Switch to a cloud model for hard queries, stay local for the easy ones.
- Privacy stance. Local by default, opt-in to cloud.
- Menu bar behaviour. Where the app lives when we are not looking at it.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price | Rating |
|---|---|---|---|---|---|
| Apple Intelligence (built-in) | Baseline system integration | macOS | Free | Free | 4.2 |
| BoltAI | Native power-user chat client | macOS | Free with limits | $34 one-time | 4.7 |
| MacGPT | Menu-bar overlay | macOS | Free tier | $29 one-time | 4.5 |
| Enchanted | Ollama front-end with FM support | macOS, iOS | Free | Free | 4.6 |
| Reor | Local-first note-taking with AI | macOS, Windows, Linux | Free | Free | 4.4 |
| MindMac | Prompt library and hotkey chat | macOS | Free trial | $30 one-time | 4.5 |
| Bezel | Screen-aware ambient assistant | macOS | Free tier | $8 per month | 4.4 |
| Elephas | Writing assistant in every app | macOS, iOS | Free trial | $59 one-time | 4.5 |
The apps
1. Apple Intelligence, best for baseline system integration
The built-in Apple Intelligence stack (Writing Tools, Summaries, Genmoji, Siri’s new answers) is the free starting point. Every text field on the system gains rewrite, proofread, and summarise. Notification summaries clean up the pile that stacks up after a break. Third-party apps can adopt the same Writing Tools UI. For most users, this is the only local AI they need.
Where it falls short: No prompt library. No custom system prompts. Long queries fall back to Private Compute in the cloud, which some users want to avoid.
Pricing:
- Free: All Apple Intelligence features.
- Paid: None.
Platforms: macOS (Sequoia and newer on Apple Silicon).
Download: Apple
Bottom line: Apple Intelligence is the free floor. Install one of the picks below on top for anything past rewrite and summarise.
2. BoltAI, best for a native power-user chat client
BoltAI is a native macOS chat client that treats Apple’s on-device Foundation Models as a first-class provider alongside GPT-5, Claude, and local Ollama models. Prompt library, hotkey overlay, inline app assistants for Xcode and Notes. The one-time thirty-four-dollar license is the cleanest pricing story in the category.
Where it falls short: Some users report the prompt library UI is buried a level too deep. Windows client is not on the roadmap.
Pricing:
- Free: Free with limits on advanced features.
- Paid: $34 one-time.
Platforms: macOS.
Download: boltai.com
Bottom line: BoltAI is the strongest all-round native chat client for macOS and it treats the on-device model as a real option, not an afterthought.
3. MacGPT, best for a menu-bar overlay
MacGPT puts a chat window in the menu bar. Hit the shortcut, ask a question, get an answer. The recent update added Foundation Models as a provider, and the free tier now covers unlimited local queries. The app is one of the older ChatGPT clients for macOS and the UI has aged well.
Where it falls short: Fewer power-user features than BoltAI. No inline app extensions.
Pricing:
- Free: Free tier includes Foundation Models.
- Paid: $29 one-time for cloud provider support.
Platforms: macOS.
Download: macgpt.com
Bottom line: MacGPT is the free menu-bar overlay if the workflow is “hit shortcut, ask, close”.
4. Enchanted, best for Ollama front-end with FM support
Enchanted is the Ollama client the community has settled on for macOS and iOS, and the 2026 releases added a switch to Apple’s Foundation Models when the network is unavailable or when the query is a simple one. Chat history syncs across devices via iCloud. Open source under the MIT license.
Where it falls short: Ollama is the primary provider. Foundation Models support was added second and shows.
Pricing:
- Free: Everything.
- Paid: None.
Platforms: macOS, iOS.
Download: Enchanted on the Mac App Store
Bottom line: Enchanted is the free open-source pick for anyone who already runs Ollama and wants Foundation Models as the fallback.
5. Reor, best for local-first note-taking with AI
Reor is a note-taking app where every note is a Markdown file on disk and the AI runs locally. Foundation Models handle the fast queries, larger local models handle the heavier retrieval. Reor’s AutoLink feature draws connections between notes automatically. Open source and cross-platform.
Where it falls short: Not a chat app. Users looking for the ChatGPT shape end up on BoltAI or MacGPT.
Pricing:
- Free: Everything.
- Paid: None.
Platforms: macOS, Windows, Linux.
Download: reorproject.org
Bottom line: Reor is the note-taking pick if we want a second brain that runs entirely on the machine.
6. MindMac, best for prompt library and hotkey chat
MindMac built a prompt library that feels like an app of its own. Save prompts by category, invoke them by hotkey, chain them together for repeatable workflows. Foundation Models integration lets the fast prompts run locally and the heavier ones fall back to a cloud provider we already own a key for.
Where it falls short: UI is denser than BoltAI. Some users bounce off the sidebar-first layout.
Pricing:
- Free: Trial.
- Paid: $29.99 one-time.
Platforms: macOS.
Download: mindmac.app
Bottom line: MindMac is the pick for anyone whose AI workflow is a saved-prompt catalog we invoke by shortcut.
7. Bezel, best for screen-aware ambient assistant
Bezel reads the active window and answers about what is on screen. Foundation Models handle the fast context reads, and the app falls back to Anthropic or OpenAI when the query needs more. Bezel is one of the closest matches on macOS to the ambient AI assistant XDA has been writing about.
Where it falls short: Subscription pricing where most competitors sell one-time licenses.
Pricing:
- Free: Tier with limited daily queries.
- Paid: $8 per month.
Platforms: macOS.
Download: bezel.it
Bottom line: Bezel is the pick if we want the assistant to see what we are looking at.
8. Elephas, best for writing assistant in every app
Elephas is a floating assistant that lives one hotkey away in every app. Foundation Models power the rewrite and summarise flows. The Super Prompt feature packages up common workflows (rewrite, tone shift, summarise, translate) into a single palette. iCloud sync between Mac and iPhone.
Where it falls short: Fifty-nine dollars is on the high end. Some users bounce off the naming (Super Prompts, Super Brain, Super Chat).
Pricing:
- Free: Trial.
- Paid: $59.99 one-time for Solo license.
Platforms: macOS, iOS.
Download: elephas.app
Bottom line: Elephas is the pick if the workflow is “every text field, one hotkey, rewrite in place”.
How to pick the right one
If we only want the built-in features: stick with Apple Intelligence. If we want a native chat client that treats Foundation Models as a real option: BoltAI. If a menu-bar overlay is enough: MacGPT. If Ollama is already installed and we want a free front-end: Enchanted. If notes come first and AI second: Reor. If saved prompts are the workflow: MindMac. If we want a screen-aware assistant: Bezel. If we want rewrite in every text field: Elephas.
FAQ
Is Apple Intelligence really free? Yes. Every Apple Silicon Mac on macOS Sequoia and newer includes the Foundation Models framework. Apps that use the on-device model do not incur cloud costs.
Do I need a subscription to use on-device AI on Mac? No. Apple’s built-in features are free, and open-source picks (Enchanted, Reor) add capability without a subscription.
Which local model does Apple use? Apple Intelligence uses a family of Foundation Models optimised for on-device inference on Apple Silicon. Third-party apps call them via Apple’s Foundation Models framework.
Can Apple Intelligence run offline? Yes. Most Writing Tools, Summaries, and Genmoji queries run entirely on device with no network needed. Private Compute and third-party providers require a network connection.
Is on-device AI on Mac private? Apple’s on-device model does not send data off the machine. Private Compute is opt-in and Apple’s stated architecture keeps data unreadable to Apple itself. Third-party apps that fall back to OpenAI or Anthropic do send data to those providers.
