Wiring a model
into your notes.
Where the keys go, what each provider is good for, what every AI feature does — and how to send one feature to a model on your own machine while the rest go elsewhere.
Four steps, and
two are optional
Open Settings → AI
Press Cmd + , (Ctrl + , on Windows and Linux) and pick the AI tab. Every key, model and provider lives there — and none of it is required.
Add a provider
“Add provider” offers Claude, OpenAI and Gemini; each becomes a card with its own key and model. Apple Intelligence and CLI agents sit underneath and need no key.
Choose a default
One provider runs every feature. Pick the one you would want for a rewrite you did not think hard about — that is what a default is for.
Override where it matters
Autocomplete gets its own dropdown, and a local model can claim individual features. Anything you leave alone keeps the default.
New here? First choose a folder for your notes, then meet the selection toolbar and slash menu that most of this hangs off.
What each one needs,
and what it costs
Claude
Your own Anthropic key, billed on your account with no markup from us. Strong on long documents and careful rewriting, and it calls tools natively — which the agent prefers.
OpenAI
Your own OpenAI key. Also the provider behind image generation, so add it if you want pictures. Tool calling is native, so the agent takes the reliable path.
Gemini
Your own Google AI key. The very large context window helps when you ask about a long note, and it can generate images too.
Apple Intelligence
On-device, no key, nothing to pay — but it needs macOS 26 or newer on Apple Silicon, switched on in System Settings. It is also the one provider with no tool support, so the agent can answer but not act there.
Local models via Ollama
No key, no network, nothing to pay, and a setup button that does the install. The honest caveat: a model on your laptop is slower and less capable than a frontier cloud one.
CLI agents
Point Ampersand at claude -p, codex or another binary. They use that tool’s own sign-in rather than an API key — but most are a local program talking to a cloud service.
Local models
and CLI agents
If Ollama is not installed, the Local models section shows one setup button that walks the whole way: open the download page, start Ollama, pull a model, and assign it across the features that suit it. Some open models can call tools and some cannot, so Ampersand asks Ollama which is which and tags them; a Tool-calling setting can force that on or off. Autocomplete is never handed to a local model on its own — inline suggestions are a latency question only your hardware can answer.
A CLI agent is added by pointing Ampersand at a command: claude -p and codex are presets, and any other binary can be registered with its own arguments. The first run shows the absolute path it resolved to and asks you to approve it — nothing is spawned before you do, no command goes through a shell, and both run time and output size are capped.
Eight things it does,
and where each one lives
Nothing here runs on its own — a request happens because you asked for one.
Autocomplete
A greyed-out continuation of the sentence you are typing; Tab accepts, Escape dismisses. Off by default, with its own provider dropdown. Free.
Selection tools
Select text and the toolbar gains an AI menu: Rewrite in five tones, Summarize, Translate, Grammar, Expand. The result replaces the selection as one undoable edit. Free.
Continue writing
Picks up from the cursor in the voice of what is already there. Run it from the command palette when you stall. Free.
Ask this document
A conversation about the note in front of you, with its full text as context. Insert an answer, or leave it in the dialog. Pro.
Ask your vault
The same dialog, but Ampersand searches every note for your question first and attaches what it finds. Encrypted notes are excluded from that search. Pro.
The agent
A tool-using assistant beside the editor. It can read the note you have open, replace an exact passage, append to it or rewrite it, and search your vault or fetch a URL you give it. No web search, no shell, no other notes — and every edit lands where you can undo it. Pro.
Image generation
Describe a picture, preview it, insert it. The file is saved into your vault with your prompt as its alt text. Needs an OpenAI or Gemini key. Pro.
Prompt library
Saved prompts that insert text, transform a selection or open a question. Selection, document and title are filled in for you; any other variable becomes a field to fill. Pro.
Default provider Claude
Autocomplete Ollama · local
Rewrite Ollama · local
Meeting summary Ollama · local
Ask Claude (default)
Agent Claude (default)
One default,
and exceptions
The default provider runs everything. Two exceptions sit on top of it: the autocomplete dropdown, and per-feature assignment for any local model you install — rewrite, summarize, ask, grammar, expand, continue writing, the agent, and the summary written after you record a meeting.
That last one is worth setting deliberately: a transcript contains what other people said, and they never chose a provider.
- A fast local model for inline suggestions
- A cloud model for rewriting and the agent
- On-device for anything sensitive
Where your words
actually go
Bring-your-own-key means what it sounds like: the request goes from your machine to the provider you chose, on your key, under that provider's terms. Ampersand runs no AI service and proxies nothing — and on Apple Intelligence or a local model, the request never leaves the computer at all.
What gets sent is the smallest thing that answers the question — a selection for a rewrite, the note for “Ask this document”, search results for “Ask your vault”. That search skips notes you have encrypted, so a locked note cannot become cloud context by accident, and running a cloud feature on one asks first.
Keys are stored on the machine you pasted them into and excluded from iCloud settings sync — preferences travel, credentials do not. And one thing to be clear-eyed about: a CLI agent is a local binary, but most sign into a cloud service of their own, so “runs on my Mac” is not the same claim as “stays on my Mac”. The specifications table sets it out line by line.
None of it is required
Free vs Pro
Good to know
Where do I put my Claude or OpenAI API key in Ampersand?
In Settings → AI. Press Cmd + , (Ctrl + , on Windows and Linux), open the AI tab, choose “Add provider”, and paste the key into that provider’s card. It is stored on that machine only and excluded from iCloud settings sync, so it never travels to another computer.
Can I use a different AI model for autocomplete than for everything else?
Yes. Autocomplete has its own provider dropdown in Settings → AI, separate from the default that runs every other feature. Local models can also claim individual features — rewrite, summarize, ask, grammar, expand, continue writing, the agent and meeting summaries — and everything else keeps the default.
Can I use Ampersand’s AI without sending my writing to the cloud?
Yes. Apple Intelligence runs on-device on a Mac with Apple Silicon and macOS 26 or newer, and local models run through Ollama on your own machine. Point the default provider — and autocomplete, if you use it — at either, and nothing you write leaves the computer.
What does the Ampersand AI agent actually change in my note?
It can read the document you have open, replace an exact passage, append to the end, or rewrite the whole body — plus search your vault and fetch a URL you give it. Every change lands in the editor as a normal edit you can watch and undo, and it never writes to another note.
Can the Ampersand agent search the web?
No. It has no search engine. It can fetch a single http or https URL you give it, or one written in the note, and read that page’s text — nothing more. If you need search results, find the page yourself and paste the link.
Which AI features in Ampersand are free and which need Pro?
Autocomplete, rewrite, summarize, translate, grammar, expand and continue writing are free forever. The agent, Ask this document, Ask your vault, image generation and the prompt library are part of Ampersand Pro, a one-time purchase. There is no AI subscription from us either way.
What happens if my local Ollama model is not running?
Ampersand checks before sending anything. If the local backend does not answer, the request falls back to your default provider and a notice says which model handled it instead, so a stopped Ollama never looks like broken AI. Start it again and the next request goes local.
Does Ampersand warn me before sending an encrypted note to a cloud model?
Yes. If the note is encrypted and the feature is routed to Claude, OpenAI or Gemini, Ampersand asks for confirmation before anything is sent, and inline autocomplete is suppressed on that note entirely. On Apple Intelligence or a local model there is no prompt, because nothing leaves the machine.
Put the model you trust
next to your notes.
Download Ampersand, add one provider, and route the rest later. The editor works perfectly with none of it configured.
Free to start · Mac, Windows & Linux · No account required