Skip to content

The assistant ​

The assistant turns a sentence into one of the commands the page already has. Open it with the sparkle button in the header or Ctrl+J.

The assistant, before it is turned on

It runs in your browser ​

The assistant is a small model that runs on your device. Turning it on downloads it once (about 35 MB). Nothing you type is sent anywhere.

It is not a chat model. It does not write prose or answer questions: it picks a command and fills in its arguments. It understands short English sentences.

What you can ask for ​

On a record: change priority to low, mark it as done, rename it to Spring launch.

The assistant proposes a change

In a list: show only urgent ones, group by status, find the invoice ticket.

Anywhere: go to members, new project called Trade fair, commit, new branch, switch to planning, show history.

Pasting a longer text (a mail, say) starts a new record with its form filled in from the text.

What it does on its own ​

  • A change to data is always shown first and runs when you click Run.
  • Going somewhere, filtering and searching run at once when the assistant is sure enough.
  • When it is not sure it asks, and when it cannot match your request it says so and offers the nearest commands.

The panel shows how sure it was ("92% sure").

What to expect ​

Measured on a fixed set of test requests, the assistant picks the right command for about three in four. Plainly worded requests (change status to done) nearly always work; looser wording (we finished this) works less often. It reads English only.

Because every change waits for Run, a wrong guess costs a click, not data.

Every picture on this site was saved by a test that had just seen the feature work.