Blog

Google published A2UI Express on 27 October. I ran it against our 557-atom catalogue on six briefs and seven models: fewer tokens on every model, the 80x was not confirmed in our tests. The bigger result is about the catalogue: keep most of a screen curated and fixed, and route each request to a small slice before the model writes.

The A2UI Atomic Catalog adds 20 D3-powered chart atoms and a morph atom that turns a treemap into a sunburst from a one-line JSON patch, all through MCP.

I audited my own AI-built component catalog and site for accessibility and turned the audit into an open-source scoring skill. The site went from 69 to 94, back to 69 when I switched on four rules my scanner had been skipping, then to 99, and to 100 on the 599 pages meant for readers under the stricter release version. The lesson: write the rules into the spec, and check what your scanner isn't checking.

A2UI now builds Google Slides decks. An agent sends a payload, an MCP tool returns a .pptx, and Google Slides opens it. Open source, with the pipe explained.

The same A2UI payload now becomes a PowerPoint file. What's inside a .pptx, how an agent builds one through MCP, and why the text keeps a margin for a missing font.

An Android app that reads any article you share, keeps it offline and draws it natively in Jetpack Compose with A2UI. Open source, prompt included.

Google's androidx.a2ui renders agent UIs in Jetpack Compose, with Material 3 versions of the 18 basic components. One library adds the whole catalog, so an agent can show real UI.

A follow-up to the Jev atom-selection pilot, with a first look at Cloudflare's Clef as a third chooser: embeddings close most of the open-pool retrieval gap (83% to 96% recall), the selector question stays genuinely unresolved on the data available, and a real trace of an apparent Haiku result turned into a permanent protocol fix rather than a headline.

A long-standing wish list item — analyse and remember articles I read, without a spreadsheet or bookmarks I'd never touch again on mobile — turned into a durable-storage MCP App, registered as a native share target, so sharing an article IS the entire interaction.

Part 1 and 2 of this series rendered the catalogue in Gemini Enterprise using GE's own native A2UI vocabulary. This is the other path: a real MCP server, wired in as a BYO-MCP connector, rendering the same Workspace and Wall Builder through the open MCP Apps spec instead — plus a tool-selection bug, an ad-hoc composed landing page, and a second MCP App (Diagram Studio) along the way.

An empirical test of Jev, a small routing model, on six UI-generation tasks -- and the follow-up that found its F1 0.91 to 0.95 gain (up from unaided) came from the hand-written shortlist, not from Jev. With no shortlist, fixing retrieval was the real lever; Jev's defensible edge, once retrieval is fixed, is cost and speed, not accuracy; and full-page composition from a one-sentence brief on an unseen domain still fails a pre-registered bar. Method, results, the correction, and a gallery of the renders.

Quick look at a pattern for using scheduled agents for self improving

I was sceptical of MCP at first — felt like marketing, felt like 'it's just an API.' Then I mapped out where transport, envelope and document model actually sit, and realised MCP Apps isn't competing with A2UI at all. It's how the same payload reaches Claude, ChatGPT and my own site without three separate builds.

Where the state actually lives across MCP, A2UI and A2A — and an afternoon of connector debugging that turned up six defects, all the same shape, including a citation for an MCP spec document that didn't exist while a real one sat one search away.

Same connector URL, same payloads, a second host — and not a line of code changed to reach it. The same airspace playbook now renders full-screen in Claude and ChatGPT alike. Getting there cost nothing; what it exposed was three latent defects that existing connections had been quietly hiding.

Add one connector URL and Claude renders real, interactive UI from the catalog's 474 typed atoms — charts, gauges, steppers, dashboards — inline in the conversation, without writing a line of HTML.

Gemini Enterprise only renders the 18 standard-catalog A2UI primitives, so the catalogue's own charts, gauges and dashboards can't be sent as native components. The fix: render them server-side to an image and deliver that through the one primitive that always works — with two real demos, and the same trick applied to Google Chat.

Google Chat's cardsV2 widget set is fixed — no custom layout, no charts beyond what Google ships. The fix: render the atom to a real image server-side and hand Chat a plain URL. This piece covers the real, self-hostable Cloud Run bundle, the exact wire-level trace of one request, and the Chat-setup gotchas Google's own docs don't lead with.

Every component in the A2UI v0.8 Basic Catalog, tested one at a time against a real Gemini Enterprise agent, plus two real working demo agents built entirely from them — and the wishlist those tests surfaced.

Frontier models are non-deterministic and expensive. My answer is artifact distillation — spend the model once, keep the output forever. Five problems nobody asked me to solve, and what each one taught me about building things that outlast the model that built them.