Nothing changes until you say so

Ask Snownotes to tighten a paragraph and it won't touch your note. It marks the change up in place, right where the text lives, and then waits for you.

That's a deliberate choice rather than a missing feature. If the assistant edits your writing the moment you ask, you have to read every result carefully to catch the times it went too far, and at that point you may as well have done it yourself.

The review bar

When the assistant changes your note, a bar appears at the bottom of the editor and the proposed version shows up in the document. Text on its way out is marked as removed, text on its way in sits next to it, and nothing has been written yet. You read it and choose Accept to apply it or Discard to keep what you had. If it's still streaming in and you've seen enough, Cancel stops it.

There's no undo step to remember afterwards, because nothing has happened to your text yet.

Word by word, where that helps

How finely the change gets marked up depends on how far it reaches.

If the edit stays inside a single paragraph, you get it word by word: only the words that actually changed are marked, and the rest of the sentence sits there untouched. That's usually what you want from "fix the grammar" or "tighten this", because the interesting part is the three words that moved.

If the edit spans several paragraphs - restructuring a list, splitting one paragraph into two - marking individual words turns into visual noise. There you get whole blocks instead, with the old ones marked as leaving and the new ones as arriving.

Asking versus editing

By default the assistant works out whether you want an answer or a change. When you'd rather be explicit, start your message with /ask for an answer that leaves the note alone, or /edit followed by an instruction for one that doesn't.

Edits act on the smallest thing you've pointed at. Select a sentence and only that sentence changes. Leave the cursor in a paragraph and it can work on that paragraph or the whole note. Point at nothing and it takes the note.

It can also do more than reply. It can search your other notes, open one to read it, and create a new note. Each of those appears as its own row in the conversation with a link you can follow, so you can see what it did rather than working it out from the answer.

Not every edit is worth a sentence of explanation. The AI button in the editor toolbar opens a menu of ready-made actions, with a heading at the top telling you whether it's about to act on your Selection or the whole Note.

Rewrite and fix

Rewrite, Fix grammar and spelling, Expand, Shorten, Clean up messy content, or turn prose into Bullets.

Change tone

Twelve voices, from Analytical and Technical to Empathetic and Humorous, keeping the meaning and changing the register.

Translate and summarize

Translate a selection, summarize a note, or pull every task out of it with Extract todos.

Suggest the metadata

Propose a title, tags, or a category. These always read the whole note rather than your selection.

The same actions are available by typing / in the AI chat, alongside commands like /find, /help, /related, and /clear. That / is a different menu from the editor's own /, which inserts blocks.

Help you didn't ask for, that you can switch off

Some of the assistance is passive. All of it proposes rather than applies, and each piece has its own off switch in Settings.

Inline suggestions

A short continuation in faint grey at your cursor. Tab takes it, Esc dismisses it, and Arrow takes one word at a time. Typing straight past it ignores it.

Cleanup on paste

Paste a whole web page and it can strip the navigation, ads, and cookie banners while keeping your words. Set it to Off, Ask, or Automatic.

Metadata chips

Once a note is long enough, a suggested title, tag, or category appears as a chip. Chips never apply themselves.

Context budget

The footer shows how much of the model's working memory your note is filling, as a green, yellow, or red dot. Very long notes get shortened to an outline so the request still goes through.

Your model, your data

None of the above says anything about whose computer the model runs on, which is deliberate.

Snownotes talks to local model servers - LM Studio, Ollama, llama.cpp, Jan, vLLM - and to cloud providers including OpenAI, Anthropic, Google Gemini, Groq, and OpenRouter. Point it at a local server and your notes never leave your machine. Point it at a cloud provider and the note text, or just the part you're editing, goes to that provider so it can answer. Your API key is stored in your browser on that device and only ever goes to the provider you picked.

Searching your notes and finding related ones happens over your own notes and doesn't involve anybody else.

Running a model on your own machine has one platform-specific catch worth knowing about, which took us a while to work out.

You can also set two models: a main one for chat and edits, and a smaller, faster one for background work like inline suggestions. Useful when the model you actually want is too slow to keep up with your typing.

What gets sent with your request

Every request carries a bit more than what you typed. Alongside your instruction we send a set of rules about how to treat the text - which characters are formatting rather than typos, which identifiers must be copied exactly, and which language the answer should come back in. Without them a model asked to fix your grammar will also quietly reformat things it doesn't recognise. We wrote up the whole set and why each one exists separately.

Try it

Open the AI tab in the inspector with /, ask it to tighten a paragraph you're not happy with, and read the markup before you decide. If you're not in the beta yet, join the waitlist.