For developers
How Jevie decides what you mean
Say one sentence and the right thing happens: Jevie finds what your words name, does the likeliest, and plans several steps when you ask for several things.
FreeMac
Use itHow it worksLimitsWhere it's kept
Use it
open hacker news and product hunt
Several places, each found, all openedsort my inbox, then remind me to go through Action
Several pieces of work, done in ordertile grid
An exact command, done at oncewhich repo are most of them from
right after a listAnswered from the rows just shown- the row under an answer (
not this? …
)The next likeliest meanings, one ↵ away summarize this
what is this screen about
, then a pauseWhat ↵ will do is the first row (↵ does this
) before you press it; when Jevie is unsure, the next two meanings are rows under it (or this
), each one ↵ awayhelp
What can be said, as you typedraw on the screen
point at the send button
The part of Jevie your words mean, as a row while you type (Ask about a highlight · like “draw on my screen”
); ↵ on it asks that part alone, Tab locks to it!git statusA command run as written, never reinterpreteddallas current time
pair my phonoe
lock it
could you lock the computer please
Said your way, typos and all: done as the tidy sentence would bedallas now
time dallas
whats the time in dallas rn
However you say a wish, the same best answer: here the clock card with your city and Dallas, never a sentence from the webai: when is the new mac coming
(or@ai …
ask ai …
)Straight to the model, nothing looked upask claude why is the sky blue
claude: …
ask codex how do I rebase
That coding agent answers it, straight away;ask my agent …
asks the one you chose.Ask Claude
is the first row fromask cl
on and ↵ takes it; with no question yet, ↵ leavesask claude
to type after- Any words
Ask Claude “chicken recipe”
sits underAsk Google
in the Mac's palette and the phone's launcher alike (your chosen agent, if not Claude); one tap or ↵ on it answers in place how do i take a screenshot on a mac
is there a way to mute my mac quickly
Explained, never done;Do it now?
under the answer does it, one press awayorder me a pepperoni pizza
call me an uber
Jevie says plainly it can't do that yet, with the nearest thing it can;try anyway
lets it work it out step by step
How it works
Jevie first finds what your words name: an open tab, a bookmark, an app, a folder, a note, a session, and your own routines before anything else. An exact command runs at once; on a close call Jevie does the likeliest and puts the next two under it, never asking which did you mean?
.
Each thing Jevie can do recognises its own sentences in code, which is fast and sure but knows only so many ways of saying them. So the same one request that weighs your words also says which part of Jevie they are about: a quick figure (the time, the weather, a sum, a timer, a die), your calendar, mail, people, notes, music, this Mac, your files, the browser, what you did before, Jevie itself, a small tool, coding agents, or the screen. When the choice falls to something only near your words - a model's answer, a plan, a settings pane that shares a word - your sentence is said once more in the words that part understands (pair my phone
), and taken only when one of its own commands is exactly that.
That second look costs about a second, and only on those sentences.
While you type, the parts of Jevie your words mean are listed as rows, found by meaning rather than by name: each part's own example sentences are compared with yours on your Mac (DigUp's EmbeddingGemma, about 20 ms), and the one or two closest are offered with the example that matched. They are offered, never taken: ↵ on the field still does what Jevie decides, and ↵ on such a row asks that part alone, said in its own words once more when its commands do not recognise yours.
Every wish has one best answer, and every way of saying it gets that one. The time in a place is the clock card (your city and the place asked, with the hours between), the weather the forecast card, a sum or a conversion the figure worked out on your Mac, the volume the Mac's own; a model's answer, a server's text or an app that shares a word never takes these. Before anything weighs your words, the manners come off a short request (could you … please
), and a word typed a letter off (volum
, clsoe
, calender
) is spelt as Jevie's own commands say it, kept only when one of them then claims the sentence outright.
A question about how to do something (how do I…
, is there a shortcut for…
) is answered, never done: the doing is offered under the answer when Jevie has it. A request nothing on a computer can do - paying, ordering, booking, a phone call or a ride, lights or a car, posting as you, anything physical - is said so in one line, with the nearest thing Jevie can do, if there is one; trying anyway waits for your yes.
Simple questions (yes or no, a choice, a score) are answered directly; questions about the world are looked up and checked before they are shown. A sentence that is several pieces of work becomes a short plan, checked before it runs, each step handing what it found to the next. This
means whatever is in focus: the card in front, the open file, the highlighted row.
Work ends one of four ways, said the same on the Mac, the phone and Windows: Done, Needs a look, Stopped or Failed. Only work that did what you asked is Done with a tick. Something that could not be done, came back empty, was not confirmed or ended in a way Jevie does not recognise is never shown as Done: it says what happened in one line and what to try next.
A plan stops at a step that did not work rather than counting it done, and talk never claims to have changed anything; asked to do something, it says nothing was changed.
Limits
- A planned step can still need you: a sign-in, a yes.
- Jevie can pick the wrong meaning; the next two are under every answer for that reason.
- What ↵ will do is shown only after a pause, for two words or more, and only when Jevie's model is reachable; it is asked at most 12 times a minute while typing, so a fast typist may see it for the last words only.
- Said again, a sentence must become one of that part's own commands exactly; when none fits, the first choice stands. With Jevie's model unreachable there is no second look: what your words name in code is offered, never a plan or a web errand.
- A word is spelt again only when nothing else explains it: a name Jevie does not know stays as typed, and a misspelling that no command claims is left for the second look.
- The rows found by meaning need DigUp's model on this Mac; without it only rows found by name are listed. A part is offered only when your words are close to one of its examples, so a rarely-said request may get none.
ai: …
answers from what the model knows and says when that may be out of date.ask <agent> …
is answered by Claude Code, Codex, or the coding agent you chose; another agent named,ask chatgpt
orask perplexity
opens its site instead.ask claude to …
is still words for a running session.- Whether something is beyond Jevie is Jevie's model's judgement, taken only when it is sure and saw nothing of Jevie's near it; anything else is planned as before.
Where it's kept
Every decision is recorded in ~/.jevie/gate.jsonl and shown on the record page (http://127.0.0.1:63798/).
In Jevie, help decide
answers from this page as you type.
Related
Updated 7 November 2026 · Something wrong or missing? Tell us