Docket

A business remembers less than it thinks it does.

The agreement is in an email. The reason for the price is in a call nobody wrote down. The number that would settle it was in a statement you filed in March.

Talk to us No sign-up. Setup is done with you, not by you.

01What it is

One place for everything your business writes down — and a way to ask for it back.

Docket takes your documents, decisions, commitments and numbers in through whichever door is nearest — a paste, a link, a file, a whole folder — and reads every one of them passage by passage.

Then you ask for it in a sentence, the way you would ask somebody who was in the room. Not a filename, not a folder, not a record number.

It runs on a Postgres database you own, on your own machine, and the AI assistant you already use can read it through one open interface.

02The problem

Your tools are good at writing things down. None of them are good at finding things again.

A point-of-sale captures every sale. A scheduler captures every job. Each one is faithful about its own slice, and each one is a locked room.

The questions that cross two tools, or reach back further than this quarter, get answered from somebody's memory or not at all.

A customer record for Cascade Ridge Property Management. A headline reads two agreements are on the record, you owe 3, they owe you 1. Below it a timeline titled How this has gone plots nine records from February to September, clustered to the right with a long empty run after February. Under that, two dated agreements about irrigation controllers and a per-visit rate hold.

What was agreed is its own dated record, attached to the person and the company. The timeline always draws today, so a relationship that went quiet shows as an empty run rather than as a cluster filling the width.

03Today

What is late, what you owe, and what you are owed.

Four zones on one screen. The fourth is the work no deadline will ever make urgent — which is the work that otherwise stays invisible forever.

The Today screen. A three-line headline reads: 2 commitments are late, 2 people owe you answers, hardscape gross margin is 12.3 points under plan. Beneath it two columns, Owed by me and Owed to me, list commitments with due dates, how many days late they are, and a line naming who is waiting.

Off-plan counts carry their denominator. Twelve of eighteen is a bad month; twelve of two hundred is a Tuesday.

04Plan vs actual

It will tell you a number is off. It will not guess whether that is good news.

Over on sales is good news. Over on labour cost is bad news, and the number alone cannot tell them apart.

So every metric carries which direction is good, set by you. One that has not been told is stated and not graded — and draws no bar at all.

The Plan vs actual screen for July 2026, headed 12 of 18 metrics are off plan. A table lists metrics with their plan, actual, and a variance bar leaning left or right of a centre axis, each row ending with the variance in dollars or points and a verdict such as behind plan or over plan.

Bars are scaled within a period, never across the whole table — one scale over both draws every percentage as a hairline.

05What is coming

The problem is rarely this week.

Sixteen weeks, one column each — four behind you and twelve ahead. Nothing is coloured good or bad; a heavy week is just a heavy week.

What you are looking for is the empty column: the room you did not know you had, to move something into.

The Schedule screen, headed: 8 commitments have no day set aside, week of 28 Sep is the heaviest ahead of you, nothing at all is landing the week of 21 Sep. A chart spanning 10 August to 30 November shows three columns of work in navy, one carrying a dark red cap for work already late, against a long run of empty weeks.

Work already late sits as a separate cap rather than as part of the week's height — a thing you are late on is a different fact from a thing you have to do.

06In, and back out

Nothing needs filing, so nothing gets lost.

You never choose a folder, and you are never asked which job a document belongs to. It is read, indexed passage by passage, and it surfaces on its own wherever it turns out to be relevant.

The Put something in screen, headed: hand it a statement, a transcript or a contract, it reads the whole thing and indexes it passage by passage. Below are four ways in — a box to read a link, buttons to choose files or a whole folder, and fields to paste a document in with an optional title. — Four ways in: a link, a file, a folder, a paste.

A single commitment, Re-bid the Franklin Avenue paver replacement, marked 21 days late and linked to a named contact. Below it a section headed What the corpus knows about this, 8 found, listing an agreement, a call transcript and a supplier price change, each with a strength ladder, a match percentage and an excerpt. — And out: one job, with everything the corpus already knew about it.

07Finding it again

Ask in the words you actually remember.

Two passes run — one by meaning, one by wording — and a re-ranking model decides which passages actually answer the question.

Weak matches are dropped rather than shown, because a search that always returns something teaches you to distrust everything it returns.

The Library screen after searching the phrase what did we agree about the irrigation controllers. Two results are returned: an agreement at 60 percent similar and a call recording transcript at 44 percent relevance, each with a five-tick strength ladder beside it and an excerpt of the matching passage.

Each hit says which instrument measured it. A re-ranked relevance and a raw similarity are not the same number, so they are never printed as one.

08Whose record it is

A database you can open, not a service holding your memory hostage.

It is Postgres. Rows in a standard database, not an opaque index. Anything that speaks SQL can read all of it, including a plain export.
Any AI can read it. Forty-five tools over the Model Context Protocol, so the assistant you already use works against your corpus — with no copy of it living anywhere else.
Your text is fenced before a model sees it. A sentence inside an uploaded document cannot issue orders to the assistant reading it.
A delete says whether it can be undone — before it acts, not after.
Links out are never fetched. A link you save is a door, not something to go and read. Nothing behind your login is quietly copied in.

09Where this is today

What is built, and what is honestly not.

Early, and run with a small number of businesses rather than sold to anybody who lands on a page.

Built

Everything pictured above, the four ways into the corpus, and the full tool interface.

Built

Importing a profit-and-loss statement — it proposes the numbers, writes nothing until you accept, and re-checks every field before it does.

Not yet

Calendar sync, jobs, inventory, deals and equipment. Each is designed and deliberately unbuilt until a real business needs it — a table nobody fills is worse than no table.

Self-serve sign-up. An install is stood up with you, on your data, which is why this page asks you to get in touch rather than to enter a card.

10Getting started

If any of this sounds like your week, say so.

The most useful first conversation is not a demo. It is you describing the last time you needed something written down six months ago and could not find it.

Talk to us A reply from a person, usually the same day.