Answers that arriveas interfaces.
A2UI is the generative interface layer. The agent does not pick a template; it composes a layout from 160+ widget types for the specific question you asked, then throws it away when you are done.
Type anything. Watch it compose.
This is the desktop's renderer, not a recreation of it: same widget set, same light theme, same card frame down to the footer titlebar. Thirty-three worked compositions with invented data; anything else still composes something, because the OS does not refuse.
How much runway do we have left?
Where the money is
Read from the ledger export on this machine. Nothing left the device.
Three months under. Compute is the only line above forecast, up 19% since the eval cluster came online.
Payroll lands where it was modelled, so the variance is real rather than timing.
Drawn by the desktop’s own renderer, from the schema the agent writes against: same widget set, same light theme, same card frame. Composed for Chat Dashboard. The data is invented; on your machine it would be your own.
And this is where you start from.
The Chat Dashboard is the front door to every agent, and it opens empty. One box, no app grid, nothing to configure. The layout above is what arrives after you ask.
Hi there,
Where should we start?
- Create image
- Create music
- Boost my day
- Write anything
Interface, rebuilt in the page · v0.1.10
This is not a chat window with better CSS.
- Returns prose you then have to act on
- Formatting is decoration, not structure
- You copy numbers into a spreadsheet yourself
- Every answer looks the same
- Returns a working interface you act inside
- Structure is the answer: charts, tables, controls
- Numbers arrive in a table you can sort and export
- Every answer is shaped like its own question
160+ widget types, composed on demand.
A sample of the library, live. Drag a slider, flip a toggle, open a section. These are the components themselves, captioned with the type name the agent writes to ask for one.
Draft
Release notes for 0.1.10.
- Wayland regression resolved
- Restore is now atomic
Three months under forecast.
How a layout gets chosen.
The composition step is deterministic enough to audit and loose enough to be useful.
Intent is typed, not guessed
The request is parsed into a structured task with a shape (comparison, extraction, monitoring, authoring) and that shape constrains which widgets are eligible.
Widgets are selected by data, not by keyword
A time series gets a line chart. A ranked set gets bars. A record set gets a table. The selection follows the returned data, so it stays right when the data changes.
Layout follows importance
Widgets are placed on a six-column grid in answer order: the thing you asked for first, supporting evidence after, controls last.
Run it on your own machine.
Free to download. Nothing leaves the device unless you connect it. Enterprise deployment is a conversation away.