Local AI or Claude for your email: when to use which
On-device AI keeps your mail on your Mac; Claude and ChatGPT go further but send what they read to the cloud. A practical guide to using both.
There are now two quite different ways to get AI help with email on a Mac, and they’re good at different things.
On-device AI runs on your Mac’s own chip: Apple Intelligence, or a small open model you download once. Nothing leaves the Mac, it works offline, and it costs nothing per question.
A cloud AI app, such as Claude, ChatGPT or Codex, reads your mail through a connection on your Mac (an email MCP server) and sends what it reads to its provider to produce an answer. It’s far more capable, and your mail leaves the Mac.
What each is good at
| Job | On your Mac | Claude or ChatGPT |
|---|---|---|
| Summarise one thread | Good | Very good |
| Draft a short reply in your tone | Good | Very good |
| “What needs me today?” | Good | Good |
| Summary of the past month | Shorter, broad strokes | Thorough, with detail |
| Compare many threads, spot patterns | Limited | Strong |
| Read several long attachments together | Limited | Strong |
| Works offline | Yes | No |
| Mail leaves your Mac | Never | What it reads goes to the provider |
| Cost per question | None | Your AI subscription |
The gap comes down to size. A small model on a Mac reads a few thousand words at a time; a large cloud model reads a whole month of mail summaries and still has room to think.
A practical way to use both
- Default to on-device for the dozens of small things in a day: a summary before you open a long thread, a first draft, “anything urgent?”.
- Reach for Claude when the question is big: “what did I miss while I was away”, “summarise the whole contract discussion”, “what do I owe each client”. Here are 30 prompts that work well.
- Keep sensitive mail local. Hide those accounts, folders or senders from agents, so only the on-device assistant ever reads them.
How Glassmail does it
Glassmail has both, and makes the line between them visible:
- Its built-in assistant runs on Apple Intelligence (macOS 26 on Apple silicon) or a downloaded Qwen or Gemma model on any Mac, Intel included. During setup it recommends the best model your Mac can run. Your mail isn’t sent anywhere. More in Private AI for email on a Mac.
- You can also connect Claude, Claude Code, Codex, Cursor and others in Settings › Agents (connect Claude step by step, or the other agent guides). Glassmail is clear about the trade: “Mail you ask about is sent to that app’s AI provider. Glassmail’s built-in AI stays on your Mac.”
- When you ask the built-in assistant something broad, like a summary of the past two weeks, and you’ve connected an AI app, it offers Ask Claude for a deeper answer. That copies a ready question that points Claude at Glassmail’s
overviewtool for the right number of days, and opens Claude.
Neither is the right answer for everything. On-device for the everyday, Claude for the big questions, and a clear line between them, is.
Quick answers
Is on-device AI good enough for email?
For quick jobs, yes. Summarising a thread, drafting a short reply, or answering "what needs me today" works well with Apple Intelligence or a small downloaded model, and nothing leaves the Mac.
When is Claude better than on-device AI for email?
For long or broad questions, such as a summary of a month of mail, comparing many threads, or reading several long attachments together. Large cloud models read more at once and reason better, at the cost of sending what they read to the AI provider.
Can I use both?
Yes. In Glassmail, the built-in assistant runs on your Mac, and you can also connect Claude, Codex or Cursor. When a question is too broad for the local model, Glassmail offers to hand it to the AI app you connected.