Can AI summarise my email without sending it to the cloud?

Yes. On a Mac, Apple Intelligence and small open models such as Gemma and Qwen can run entirely on the machine. An email app built for them can summarise threads, draft replies and answer questions about your mail without sending the content anywhere.

Which Macs can run on-device AI for email?

Apple Intelligence needs an Apple silicon Mac with Apple Intelligence turned on; Glassmail uses it on macOS 26 or later. On other Macs, including Intel Macs, a small model downloaded once can run with llama.cpp, more slowly.

Is on-device AI as good as ChatGPT or Claude for email?

For short, everyday tasks such as summarising a thread or drafting a two-paragraph reply, small on-device models are good. For long, complex writing, the large cloud models are better. On-device AI trades some quality for privacy and offline use.

Private AI for email on a Mac: what stays on your Mac, and what doesn't

AI that summarises and writes email usually sends your inbox to the cloud. Here's how on-device AI works on a Mac, what it needs, and what it can't do.

AI that summarises a thread or drafts a reply is genuinely useful. But in most email apps, “AI” means your messages are sent to a server, sometimes the app’s and sometimes an AI company’s, to be processed. For a lot of people, and a lot of jobs, that’s a non-starter.

There’s another way: run the model on the Mac itself.

Cloud AI vs on-device AI

Cloud AI sends the text of your emails to a large model running in a data centre. It’s powerful, but your mail leaves your Mac, often under terms you haven’t read, and it stops working offline.

On-device AI runs a smaller model on your Mac’s own chip. Nothing is sent anywhere: the email, the prompt and the answer all stay on the Mac. It works on a plane, there’s no per-use cost, and there’s no account to create.

What it can do well

Small on-device models are good at the everyday jobs that make up most of email:

Summarising a thread into a few lines: what it’s about, what you’re being asked, and by when.
Drafting a short reply from a hint like “yes, but next week”, in your usual tone.
Changing a draft: shorter, friendlier, more formal, or just fixing spelling.
Answering questions about your inbox, like “what needs me today?”, by pointing at the actual conversations.

What it can’t do (yet)

Long, nuanced writing. A small model writes a solid two-paragraph reply; it won’t write your board update as well as a large cloud model.
Facts it wasn’t given. Like every model, it can be wrong. A good app makes it easy to check, and never sends anything on its own.
Very long threads at once. Small models read less at a time, so apps summarise long conversations in pieces.

What your Mac needs

Apple Intelligence runs on Apple silicon Macs with it turned on in System Settings. Apps can use Apple’s on-device model directly; Glassmail does on macOS 26 or later.
Open models such as Google’s Gemma or Alibaba’s Qwen can run on any recent Mac, including Intel Macs and older macOS versions, through llama.cpp. You download the model once (roughly 1–3 GB), then it runs offline.

How Glassmail does it

Glassmail uses Apple Intelligence when your Mac has it, and otherwise lets you download a small model in Settings › AI. The assistant, thread summaries and the AI writer all run on your Mac. Every draft is marked as AI-written until you change it, every change can be undone, and nothing is ever sent without you pressing Send.

One honest caveat: if you choose to connect an outside AI agent, like Claude or Cursor, to search your mail, that agent may send what it reads to its own servers. That’s off until you turn it on, and our privacy policy lists every connection the app makes.