AI

AI in Nifty is off until you add a model and turn on a feature. AI can give a short title and a type to notes that don't have them, pick out what's worth remembering in your notes, read the images you add, and summarise your documents. You choose the model, and it can run on your own computer or server.

What AI does

Note titles. When you write a note without a title, Nifty asks your model for one about 10 minutes after you stop editing. The title has at most eight words and is in the note's language. As the note grows, Nifty keeps the title up to date.

• A title that AI wrote has a sparkle beside it in lists, and Title by AI appears above the note.
• If you write, change or clear a title yourself, AI won't change that note's title again.
• Very short notes, under 20 characters, aren't given a title or a type.

The same feature chooses a type for a note that doesn't have one, such as a meeting or an idea, including a note whose title you wrote. As the note grows, AI can change a type it chose. AI never changes a type you chose or one a calendar event gave.

Note insights. About 10 minutes after you stop editing a note, Nifty asks your model to pick out what's worth remembering in it: plans, promises you made and promises made to you, what needs to happen next, decisions, dates, worries, questions and news about the people and projects it mentions. Nifty also reads the note again when you link or unlink a task or an event.

• Each insight is based on words in the note. If the model offers one that the note doesn't say, Nifty leaves it out.
• Nifty matches each insight to the people, projects, tasks and companies in Nifty that it's about. If a note @mentions someone or uses their full name, their first name anywhere in the note means them, even when other people share it. Otherwise, a first name that several people share, such as "Sam", stays unmatched rather than guessed.
• A name straight after another person's name or "his", "her" or "their", such as "Simon's Dad" or "her Mum", means that person's relative. If your Dad has the alias "Dad", "Dad" and "my Dad" match him, but "Simon's Dad" doesn't.
• An insight that's already a task says so, including a task the note names by its title.
• When an insight is about something to come, Nifty works out its date from phrases such as "next week", "by Friday" or "3 to 7 November".
• Nifty keeps an insight that has a date even when it can't tell who it's about, such as a plan for someone who isn't in Nifty yet.
• A note of 150 words or more also gets a summary above its text. A longer note's summary has more points, up to 12.
• To see everything Nifty found, choose Insights in the note's ⋯ menu, or See all … insights under its summary. To read a note Nifty hasn't read yet straight away, choose Generate insights in its ⋯ menu. Nifty reads it and opens Insights. Choose Show the words to see where each insight came from. Choose Read again to have Nifty read the note now.
• To remove a note's summary, choose Remove summary… in the note's ⋯ menu. Nifty won't write a new one when you edit the note. To get a summary back, choose Read again in Insights.
• Insights change nothing in your note, and Nifty doesn't make tasks or facts from them.

Correct an insight

In Insights, you can put right what the model got wrong. Nifty keeps your corrections when it reads the note again.

• Choose Not right to mark an insight that's wrong. Nifty won't use it anywhere else.
• To unlink a person, project, task or company the insight isn't about, choose × on its chip.
• To link the insight to something else, choose Link… in its ⋯ menu, then find a person, project, task or company.
• To fix the wording, choose Edit… in its ⋯ menu.

Document summaries. When you add a PDF or an image, Nifty asks your model to summarise it and say what kind of document it is, such as an energy bill, a payslip, a bank statement or an insurance policy. The model also picks out its details, such as a bill's readings and charges. The summary and details appear beside the document in the viewer.

• The model reads the document's text, so scans without text and images without words are skipped.
• An image's summary covers what its text says, not only what the image is. A screenshot of a conversation is summarised as the conversation: who took part, what was asked or agreed, and any dates or amounts.
• A long PDF, or one too big for your model, is summarised from its first pages. The summary says when this happens.
• AI can make mistakes. Check the document before you rely on a figure.

Image processing. Nifty reads the text in every image itself. With this on, Nifty also sends each image you add to your model, which reads its text more reliably, says what it shows, and gives a screenshot or photo still called by its file's name, such as "Screenshot 2026-10-17 at 10.42", a better one. A name you gave it is never changed.

• Choose a model that accepts images, such as a vision model in Ollama or LM Studio, or one from OpenAI or Anthropic.
• The image itself is sent to the model, as well as its text.
• If the model can't read an image, Nifty uses the text it read itself, and the viewer says why.

Add a provider

A provider runs the models. To keep everything on your own computer or server, use Ollama.

1. Open Settings → Providers and models and choose New provider (N). If you have no providers yet, you can choose Add Ollama to fill in the details for you.
2. Give it a Name and choose its Type. Choose OpenAI-compatible for OpenAI, Ollama, LM Studio, llama.cpp and vLLM, or choose Anthropic.
3. Enter its URL, including the version, such as http://localhost:11434/v1. Anthropic's URL is filled in for you.
4. Add an API key if the provider needs one. OpenAI and Anthropic need a key, but Ollama and LM Studio don't.
5. If the provider runs on a slow computer, choose a longer time under Wait for an answer. The Standard wait is 90 seconds, or 5 minutes for documents, images and notes.
6. Choose Test connection, then Save.

Add models

1. Choose the provider's tab under Providers, then Add models.
2. Tick the models it lists, or type one under Add by name exactly as the provider names it, such as llama3.2:3b.
3. Choose the button that counts them, such as Add 2 models.

A model's ⋯ menu has Make default, Rename… and Remove….

Turn on a feature

1. In Settings → Providers and models, under Default model, choose the model that AI features use.
2. Open Settings → AI features. Turn on Note titles, Note insights, Document summaries or another feature with its switch.
3. To catch up on the notes and documents you already have, choose the action in the feature's ⋯ menu, such as Add titles and types to … notes or Read … notes. Each result appears when it's ready.

When a feature is on, you can choose Options… in its ⋯ menu. There, you can give the feature its own Model instead of the default, and set Reasoning and Max reply tokens. Turn reasoning off for quicker answers.

Until a provider has a model, you can see every feature, but you can't turn any of them on.

If a provider stops answering, its tab in Providers and models shows an alert icon. The tab says when the last request failed and why. When the provider is working again, choose Check again. If it says the provider didn't answer in time, choose Edit… in the provider's ⋯ menu and choose a longer time under Wait for an answer.

Nifty asks a provider one thing at a time, so while it waits for one answer, other AI features wait their turn.

Privacy

• Nothing is sent anywhere until you add a provider and turn on a feature.
• For a title or a type, Nifty sends up to 8,000 characters of the note's text to the model you chose. People, tags and projects are sent by name, and images aren't sent. A note whose title you wrote is sent too, unless you or a calendar event gave it a type.
• For a note's insights and summary, Nifty sends up to 24,000 characters of the note's text, with its title, type, the day it was written and your name. People, tags and projects are sent by name, and images aren't sent. Nifty also sends the people, projects, tasks and companies the note may be about: people's job titles, the status of projects and tasks, tasks' due dates and who they're delegated to, and the note's event with its date and attendees.
• For a summary, Nifty sends up to 60,000 characters of the document's text, along with its name, your name, today's date and your time zone. These help the model tell what's yours and read dates the way you write them. For an image read by Image processing, its description is sent too.
• For Image processing, Nifty sends the image itself, at most 1568 pixels across, with its name and up to 20,000 characters of the text Nifty read from it.
• With a local provider such as Ollama, your notes never leave your computers. With an online provider, such as OpenAI or Anthropic, the text goes to that company under its terms.
• API keys are stored encrypted on your server, and Nifty never shows them again after you save them.