The planner that learns how long your work actually takes.

Three things it does differently

Every sitting says why it is where it is.

Not “the AI decided”. Each hour the plan books stores the factors it was placed from and says them back beside the sitting: “Placed at the first free time after the 9:30–10:00 meeting and its 10-minute padding.”

An overcommitted day reads 164%. The number is never capped.

Work that does not fit is listed with the reason it found no place, never dropped, never squeezed in. Assumed hours say “assumed” until you confirm them.

Plans hold still.

A sitting within ninety minutes never reshuffles without a real conflict, and a repair moves the minimum. When a meeting lands on a planned morning, only the hour it displaced moves, and the plan says what moved and why.

A day, as it plans one

The close is a two-minute conversation

Every evening the day is judged against the plan it started with, from the record — the timer, the tracker, what you marked done — and said in one line, which the next morning's door repeats:

Plan kept today: 3 of 5 sittings (2h 40m of 4h 30m) — 1 moved by you, 1 missed.

Then the day's open loops answered for, what the next working day holds first, and a note read for commitments. No mood check-in, no score, no streak. What you answer feeds the learning; what you skip is still on the record as time the tracker saw with no task named.

Every week it knows your work better — here is the founder's own record

The one chart this claim rests on, live from the account this was built in. Each week: what the work took over what was sized for it — the first guess, and the plan's sizing of that guess with what the learning knew when the week began, replayed from the record. Aggregate ratios only; no task is named.

Privacy, in the policy's own words

• The desktop app records which application has focus and how long — never a window title, never anything you typed or read. Site names only if you switch them on, and then the domain alone.
• It never modifies or deletes your own calendar events. It reads them, and writes only to a separate calendar it creates for you; there is a mode where no meeting title is ever stored.
• Models read messy text and write summaries. Placement, priority and repair are plain deterministic code. Your browsing domains are never sent to any model at all.
• Settings has a delete button, and it is not a request: everything the app stores goes at once, and the sign-in record is removed by hand, which the policy says plainly. No advertising trackers, no analytics scripts.

The whole of it, written from the data model rather than a template: the privacy policy.

What it will cost

One plan, billed yearly

$16 a month

Month to month

$20 a month

Founding price, the first 100 people — held for as long as you stay

$12 a month, yearly

Trial

30 days, no card

The models cost real money per account, so every account has a spending ceiling, stated in the terms; the price is set so a heavy month stays inside it. Cancelling will be as plain as deleting. Nothing can be bought yet — the beta is closed, and the first invitations go to the waitlist.