Skip to content

AI & intelligence

Where each AI action lives, Review Links, New Note from a Prompt, typing suggestions, the Assistant, and choosing Local (Apple), Cloud (Apple) or an MLX model.

Where it runs

Every AI feature runs on Apple’s Foundation Models, on a model named for where it runs:

  • Local (Apple) — Apple’s model on your Mac, iPhone or iPad, the one Apple Intelligence runs. The default.
  • Cloud (Apple) — Apple’s server model, on Apple’s Private Cloud Compute, in the App Store version.
  • Local (model name) — an open model you download and run on your device with MLX.

On a model on your device, your notes are processed there and are not sent to us or to any AI service. On Cloud (Apple), what a feature works on — your question, and the notes it reads to answer it — is sent to Apple’s Private Cloud Compute to be processed; it isn’t stored, and it isn’t accessible to Apple. There are no third-party AI services and no API keys. Each Assistant reply, summary, suggested tag or link, rewrite and new note’s draft says which model made it, in small type beside it: Local (Apple) with a lock, say, or Cloud (Apple) with a cloud. Choosing a model says how to pick one.

Whichever model runs, a web search leaves the device when the Assistant or Research looks something up: the search terms go to DuckDuckGo — and the model writes them after reading your question and, often, your notes — and a page it then reads is fetched from that page’s site. Downloading an MLX model fetches its files from Hugging Face.

The AI features are available when the model you have chosen can run. Local (Apple) needs a device that supports Apple Intelligence, with Apple Intelligence turned on — in System Settings on the Mac, Settings on iPhone and iPad. Cloud (Apple) needs an internet connection, and not every device can use it. An MLX model needs to be downloaded, or its folder chosen. Otherwise the AI commands are unavailable, and the AI settings say why.

Everything else in HelloNotes works without them — the AI is an addition, never a requirement.

Note actions

These live in the Note menu, beside Rename and Duplicate — not in a panel of their own — and in Note Actions, the menu at the right of the bar over the open note, under the name of the model that will do the work. On iPhone they are also gathered in the AI tab. A suggestion appears in the right panel’s view that already holds that kind of information, so you find it where you would have gone looking anyway.

  • Summarise Note — a short summary of the current note, shown at the top of the Summary & Outline view. A summary is an outline at a different resolution. Save to Properties keeps it in the note’s summary: property.
  • Suggest Tags — proposes tags based on the content, in the Tags view; you choose which to keep.
  • Suggest Links — finds notes that are plausibly related, in the Links view beside the links the note already has.
  • Rewrite or Expand Note… — improve the writing, fix the grammar, make it concise, elaborate, simplify, turn it into bullet points or expand it — or describe the rewrite yourself. You see the result and choose Replace Note or Insert Below; nothing changes until you do.

Selecting text adds the actions that need your notes to make sense to the editor’s own selection menu — a right-click on the Mac, the standard selection menu on iPhone and iPad: Rewrite with AI…, Link to “note” when the phrase names a note you have, Find Related and Ask the Assistant. They sit beside Writing Tools, which the system already puts in every text view.

Can’t remember where a command lives? ⇧⌘P — File ▸ Command Palette… — opens a command palette: type a command’s name to run it.

Note ▸ Review Links… (⇧⌘L) walks the note phrase by phrase the way a spell check does — Link, Skip or Never. You see the phrase in its own sentence and the opening lines of the note it would point at, because the same phrase can deserve a link in one paragraph and not the next.

  • Every phrase that names another note is a candidate. When a note turns up more than ten, the ten whose targets are most related to what you are writing are the ones shown — a long note can otherwise produce dozens, and a list nobody finishes reading is a list that gets skipped.
  • Never is remembered for that collection, so a phrase you have rejected stops coming back. It is kept on your device rather than in your notes — a decision you made is not something to commit to a shared repository.
  • Nothing is written until you finish.

This one needs no model at all: it is an exact scan of your own text.

New note from a prompt

File ▸ New Note from a Prompt… (⌃⌘N) — also under the + in the bar over the note — writes a new note, in two modes. Write composes one from a description. Research investigates a question on the web across several angles and lands the answer as a note, with its sources gathered at the end and a record of the question it came from.

Research needs a model with room to read web pages — at least 8,000 tokens. MLX models have that, so does Cloud (Apple), and so does Local (Apple) where Apple Intelligence gives it 8,192 tokens. Research uses the Default model in Settings ▸ AI, and if that one can’t hold enough, the sheet says so rather than starting a run that cannot finish.

Either way the draft is connected to what you already have: notes you own on the topic are offered as link targets — Research can search and read them as well — and every link that comes back is checked against your collection. Anything naming a note that does not exist is turned back into ordinary text, and the draft says so. A link that looks real but points nowhere is worse than no link.

You read the whole draft — title and body, both editable — before anything is created. Cancel leaves nothing behind. The button is Create Note.

Suggest as I type

Grey text appears after the cursor when you pause at the end of a line in the Edit view. On the Mac, ⌥⇥ or → accepts it and Esc dismisses it; ⌥Esc or F5 asks for one without waiting.

The suggestion is not part of your note until you accept it. It is drawn on screen and nothing more — it cannot reach a save, a search result, your links or a Git diff.

On iPhone and iPad, tap the suggestion to accept it — or press ⌥⇥ with a keyboard attached; Esc dismisses it. Suggestions come from the Default model in Settings ▸ AI, when it runs on your device — a round trip to Apple’s servers can’t keep up with typing, so they are off while Cloud (Apple) is the default. They are off until you turn on Suggest as I type, in the Inline Completion section of the AI settings.

The Assistant

The Assistant answers questions from your notes, and shows the notes it read. It is most useful for questions you know you have written about but cannot place — “what did I decide about the pricing model?” rather than general knowledge. Open it from Assistant at the top of the sidebar, or with View ▸ Assistant (⇧⌘A); on iPhone it is in the Notes and AI tabs. It opens as a tab beside your notes rather than a window, so a note and the Assistant are a tab apart. Select some text and choose Ask the Assistant to ask it about that passage: its tab opens, and the question is asked at once — or, while a reply is still coming, put in the message field to send when it has.

With Use my notes on — the switch at the top of its tab — it searches your notes: the words in them, any section of a long note, and to the end of it. It reads what it finds, and can look things up on the web when a question needs it, saying what it is doing as it works — searching your notes, reading a note, searching the web. With the switch off it answers from the model alone and reads nothing. On a model that can’t use tools the Assistant can only chat, and says so.

The menu at the top of the tab picks the model for this conversation. It starts on the Default model from Settings, and a model picked in the menu is for this conversation only: New conversation, beside it, starts again on the default. Each reply says which model made it. On Cloud (Apple), a line above the message field says that each message sends the conversation so far to Apple’s Private Cloud Compute, with what the Assistant has read from your notes — and when today’s allowance is nearly used, it says that too; once it is used up, Local (Apple) answers instead. When a conversation grows longer than the model can hold, its earlier messages are summarised so it can go on.

It is experimental. It reads a few notes at a time, so it is not a full chat assistant, and it doesn’t change your notes: it can’t create, change or delete one. Copy what you want from its reply into a note yourself.

Skills hold reusable instructions for repeated jobs — a note named SKILL.md, with a name: and a description: in its front matter — and the Assistant loads the one a request matches. On a model with a window of at least 16,000 tokens — an MLX model on a Mac, or Cloud (Apple) in the App Store version — it can also read whole web pages and run deep research, a longer multi-step investigation across your notes and the web. With a smaller window it searches the web and works from the results.

Reporting a response

What the AI writes stays with you — in your notes, and the Assistant’s conversations in HelloNotes’ own storage on your device. Nothing is posted anywhere, and there is no one else to see it. If something it wrote is wrong or offensive, Report… is beside its answers and suggestions: an Assistant reply, a rewrite, a new note’s draft, a summary, and suggested tags or links. It opens your mail app with a message to us holding the response, what you asked, the model that made it and the app’s version; you read it, and can change or delete any of it, before you send. HelloNotes sends nothing itself.

Anything else it wrote — a suggestion shown as grey text, say — is reported by writing to us from Settings ▸ Support ▸ Contact, which anyone can use. Local (Apple) applies Apple’s safeguards, and Cloud (Apple) is Apple’s own model on Apple’s servers; an MLX model is the one you chose, and HelloNotes does not filter what it writes.

Choosing a model

Settings ▸ AI has the Default model: the one everything the app does with AI uses — Summarise, Suggest Tags and Links, Rewrite, Compose, the research behind New Note from a Prompt, the suggestions as you type (which need it to run on your device) and the Assistant, whose menu can choose another model for one conversation. It can be:

  • Local (Apple) — Apple’s model on your device, whichever one Apple Intelligence runs on your hardware. The default.
  • Cloud (Apple) — Apple’s server model, in the App Store version. It runs on Apple’s Private Cloud Compute: what a feature works on — your question, and the notes it reads to answer it — is sent there to be processed; it isn’t stored, and it isn’t accessible to Apple. It needs an internet connection, and not every device can use it; it reads far more at once than Local (Apple), and can’t make suggestions as you type. Apple sets a daily allowance. Settings and the Assistant say when you are close to it, and once it is used up Local (Apple) does the work instead until it resets — there is nothing to press, and each result still says which model made it.
  • Local (model name) — an open model from your models folder, run on your device with MLX. Every model in the folder that HelloNotes can run is listed here, one entry each, so choosing a model and choosing MLX are the same act. A folder with nothing runnable in it leaves the picker offering Apple’s models — and, if an MLX model was your choice, an entry saying the folder has no model.

Under the picker is what the model is used for, one line saying where your text goes, one saying what a web search sends — for a model that can use tools — and, if it can’t run right now, why not. With Cloud (Apple) as the default, a line there also says when today’s allowance is nearly used, or used up.

Under Assistant are the model’s own settings, as the Foundation Models framework has them and no further: Creativity (0–2) and Sampling — Automatic, Greedy, Top-k (which adds Words to sample from) or Top-p (which adds Probability), the last two also offering a Seed so a run can be repeated — plus a Maximum reply length and, on a model that reasons, Thinking. They apply to the Assistant; each writing tool asks for what its own task needs.

This is the AI tab of Settings, the same on every device. Settings opens with ⌘, or the gear in the bar over the note; on iPhone, from the ⋯ in the Notes tab. It is the only AI settings screen: every AI Settings… — at the foot of the Assistant’s model menu, for one — opens Settings on this page.

Open models with MLX

HelloNotes suggests no models. Choosing one is the advanced end of the app: a recommended list is a generation out of date within months, and whether a model is any good at your work is between you and the model.

MLX models live in one place, the models folder. Until you say otherwise that is HelloNotes’ own storage.

On a Mac your models are probably already here, in the Hugging Face cache ~/.cache/huggingface/hub that mlx_lm and the other MLX tools share — but macOS keeps HelloNotes out of that folder until you allow it. Access MLX Models… opens it, and one click on Allow allows it for good: everything in it HelloNotes can run is listed, and downloads join it rather than making a second copy of a multi-gigabyte model. Change… points the folder somewhere else afterwards. On iPhone and iPad the button is Choose Folder…, for a folder in Files or on an external drive.

  • Use a model in the folder, or Remove… it. Removing deletes it from the folder, for every tool that shares it, so HelloNotes asks first.
  • Download an MLX language model from Hugging Face by typing its name in the form organisation/model. It goes into the models folder, and a download can be cancelled. The settings link through to the mlx-community organisation, which converts most models to MLX.
  • If Hugging Face isn’t reachable where you are, put a model into the models folder another way. Once a model is on your device it works offline.

The Assistant’s menu lists every model in the folder too, so a conversation can run a different one from the default; each model in use is loaded into memory, so two at once need room for both. Pick it in the Default model picker, or with Use beside it in the models list — they are the same choice.

Not every open model can use tools — Gemma 3, for one, has no way to be shown them. With such a model the AI settings say so, the Assistant chats without reading your notes, and Research isn’t available; the writing tools work as usual. HelloNotes also says when a model’s weights are large for the device it is on.

How much a model can read

Every model has a window — how much text it can hold at once. HelloNotes asks Apple Intelligence how large the window of Local (Apple) is rather than assuming one, and asks Cloud (Apple) the same — 32,768 tokens, last time it was asked. An MLX model is given 16,384 tokens on a Mac and 8,192 on iPhone and iPad, because a larger window costs memory. Text in Chinese, Japanese or Korean takes about a token a character, so the same window holds fewer characters of it.

  • A long note is summarised whole — in parts, then the parts — rather than by its opening. One too long even for that is refused, not cut short.
  • Rewrite asks for a shorter passage if a selection is more than the model can rewrite in one go, instead of rewriting only the part that fits and replacing all of it.
  • Tags and links read a note’s opening if it is very long — they are suggestions — and tags come back in the note’s own language.
  • A long conversation goes on — when the Assistant’s conversation outgrows the window, its earlier messages are summarised rather than dropped.