The assistant

It builds the first project. Then it gets out of the way.

Nothing watches you while you work — no suggestions in the editor, no model reading your scripts, nothing sent anywhere. There is exactly one place where a model earns its keep, and it is the emptiest moment there is: a project with nothing in it. Describe the business in a few sentences and out comes a data model, windows, reports and a first set of scripts — a starting point you rename, rearrange and throw half of away.

Three rounds, and each one puts something on the table.

Not one prompt and a shrug. The data model first, because everything rests on it; then the windows and reports; then the scripts — and the third round is told the real object names, because by then the windows exist.

The Describe your business window with a few sentences about a bike repair shop typed into it
What you type. A few sentences, in your own language. The tables and fields come back named the way you wrote them — and the window says outright where the description goes.
The same window, empty, with the worked example unfolded — a description written in German
In your own language, and it means it. The example under the box is German, and that is the demonstration: describe a workshop in German and the tables come back called Kunden and Auftraege.
Round one: six tables with their fields and types, and two questions the model could not decide
Round one — the data model, plus the questions it cannot answer. Six tables with every field typed, and underneath the honest part: a few things it cannot guess. Is a status free text or a fixed list? Does the hourly rate belong to the job, the shop or the customer? You answer in one click and it works them in.
Round two: the same model, now with ten windows and five reports listed
Round two — the windows and the reports. Ten windows and five reports, each named and listed with what it shows. Still nothing written to disk: this is a plan you can still walk away from.
Round three: the reports with their fields and grouping, and fifteen scripts by path
Round three — the scripts, by the path they will live at. Fifteen of them, each addressed the way the project addresses it — and the sentence underneath is the one that matters: “Everything here is a starting point — rename it, throw half of it away, add what it missed. That is the part no model does for you.”

And then it is a project like any other.

No special mode, no “generated” badge, nothing that has to stay the way it came out. Press ▶︎ and the application runs — 41 seconds after the description was typed.

The generated application running: a start window listing its five list windows
It runs. The start window it wrote, with a button per list — and it repeats the warning rather than boasting: a starting point, not an answer.

What it costs, and what leaves your Mac.

Your description goes to Anthropic — and it says so

That one paragraph is what is sent, three times, once per round. Not your schema, not your scripts, not your data. The line stands in the window itself, in orange, where you type — not in a policy nobody opens.

Your key, in the keychain

You bring your own Anthropic key. It lives in this Mac's keychain — never in a project, never in a file, and never inside anything you build. A project travels through Git; a key must not.

A fraction of a cent — and you pick the model

Three rounds of text. Put a few euros of credit on the account and you will lose interest in the counter long before the credit runs out. There is no subscription of ours in the way, and nothing is resold. The settings offer Opus when the business is unusual, Sonnet as the balance it starts on, and Haiku when quick and cheap will do — or type the name of any model your account can reach.

Nothing is accepted on trust

Every script the model proposes goes through three gates before it is written: it has to parse, every table, field and object it names has to exist, and every window it opens has to be one the project actually has. What fails a gate is dropped, and the window says what was dropped.

The whole conversation is kept

Every prompt and every answer, verbatim, in a transcript beside the project — so a thin result can be argued about with the evidence in hand rather than from memory.

And then it stops

That is the whole of it. No assistant in the code editor, no “explain this script”, no model in a built application unless you put one there — and if you do, it is Apple's on-device model, which costs nothing and asks nobody.

The Assistant page in Neuridion's settings: provider, model, key, and what it is for
One page of settings, and one honest sentence in it. Provider, model, key — and the sentence about where the key is kept. Without a key the other two doors into a new project work exactly as before.
Why there is no assistant while you work. The editor already knows your schema — it completes your tables and fields, underlines a name that does not exist, and jumps to a definition on ⌘-click. That is knowledge, not guessing, and it is right every time. A model guessing on top of it would be right most of the time, which in a tool you use for eight hours is a different thing entirely.