Nearly all of it is a chat window. Very little of it reads your Mac.
We make one of the tools below. Brownie is ours, so read this the way you would read any comparison written by someone with a stake in it. We have tried to make the table something you could check yourself in an afternoon: every tool links to its own site, the claims are the structural ones rather than the ones that change with a pricing page, and where we could not verify something we have said so instead of guessing.
Two different products share one search term.
Model runners
You open an app, pick a model, and type at it. The model runs on your GPU, so nothing goes to a server — but it only knows what you paste into the box. Ollama, LM Studio and Jan are all this, and they are good at it.
Readers
Something that already has access to your messages, mail, notes and files, and does work on them without being asked each time. Far rarer. Apple Intelligence does a narrow version of it; Brownie does a broad one.
The distinction matters because the privacy question is completely different for each. A model runner is private because you never give it anything. A reader is handling your actual life, so the question becomes what it does with it — which is the only reason the last column of the table exists.
What each one actually is.
| Tool | What it is | Works offline | Reads what is already on your Mac | Licence & price |
|---|---|---|---|---|
| Ollama | A command-line runner for open models | Yes | No — you pipe things to it | Free, open source (MIT) |
| LM Studio | A desktop app for downloading and chatting with models, with MLX support on Apple silicon | Yes | No — chat only | Free to use, closed source |
| Jan | An open-source desktop chat app with a bundled engine and model hub | Yes | No — chat only | Free, open source (AGPL-3.0) |
| Elephas | A Mac writing and knowledge assistant, with an offline mode using local models | Offline mode available | Documents and notes you add to it | Paid — see their pricing page |
| Apple Intelligence | Built into macOS: Writing Tools, notification and Mail summaries, Siri | Partly — larger requests go to Private Cloud Compute | Yes, but in the moment: it summarises, it does not keep notes | Included with macOS |
| Brownie | A background reader: overnight it reads what is new and writes an Obsidian vault, then leaves morning cards | Yes — “This Mac only” needs no key and sends nothing | Yes: iMessage, WhatsApp, Telegram, IMAP mail, Apple Notes, Calendar, files | Free, open source (AGPL-3.0) |
Checked 1 November 2026. Prices and tiers move, so the price column links out rather than quoting a number we would not keep current — Elephas in particular is listed at several different prices by several different review sites, none of them official. Correct something by opening an issue.
Which one you want depends on one question.
It reads at night and leaves the morning ready.
At an hour you choose, your Mac wakes itself and a four-billion-parameter model on your own GPU reads everything new since the last run — messages, mail, notes, files. It writes what matters into plain Markdown notes, one per person, group and plan: an Obsidian vault on your disk, yours to keep. In the morning there are a handful of cards: a reply someone is waiting on, a promise you made, a renewal worth cancelling, each with its evidence and a draft in your voice.
Raw messages never leave the machine. Only short summaries go to whichever model you chose, and the app shows you every byte that left, kept for thirty days. Pick “This Mac only” and the figure is zero: no key, no account, nothing out. It never sends anything on your behalf — it fills in the draft and stops before Send.
If you keep an Obsidian vault, the same split shows up inside that world too, and we wrote it down: an AI assistant for Obsidian — every plugin reads the vault, almost nothing fills it. And on the one thing those notes are for: what did I promise? The Apple Intelligence row above gets its own page: finding, versus remembering.
It is free, AGPL-3.0, and in the open, so the claims above are checkable rather than promised. It needs an Apple-silicon Mac on macOS 14 or later, 8 GB of memory, and a 3.7 GB model download once.
The ones people actually ask.
