a2uicatalog

# A2UI Atomic Catalog **Real rendered UI in the conversation — charts, gauges, dashboards, study apps — without the model writing a single line of HTML.** The agent names a component from a fixed vocabulary of 474 typed atoms. A renderer that already knows that component draws it. So a hallucinated atom is a *parse error*, not a broken screen — the failure surfaces before anything reaches the user. No signup, no API key, no configuration. Add the URL and start. ``` https://a2uicatalog.ai/mcp ``` ## Try it > Using the A2UI catalog, render a stat_card with value "1,234", label "Daily users", delta "+12%". > Build me a full-screen revision app for the French Brevet — subject tabs for Maths, > Français and Histoire-Géo, with flashcards and a timed drill in each. > Show a live ATC radar for Toulouse with simulated traffic and the real LFBO weather. The second one is a single tool call: `emit_runbook_surface` stamps your content through a pre-authored composition, so the agent supplies content and makes zero layout decisions. The third pulls live METAR through a declared data proxy. ## One payload, many surfaces Every atom declares where it works, where it degrades, and where it genuinely cannot go — and those declarations are **enforced**, not documentation: | | | |---|---| | `works_on` | renders fully | | `degraded_on` | renders with something lost — an animation frozen, an ordered list instead of a stepper | | `incompatible_on` | cannot do what it is designed to do there, and says so instead of pretending | A `quiz_set` you cannot answer is not a degraded quiz, it is not a quiz. The catalogue refuses rather than shipping something that looks right and isn't. Surfaces: **web · MCP Apps · Google Apps Script · Google Meet · Google Chat · PDF · email** ## Interactive UI, not screenshots Implements **MCP Apps (SEP-1865)** — `ui://` resources over the standard `ui/*` JSON-RPC bridge. Verified rendering full-screen in **both Claude and ChatGPT** from an identical payload, which is the point: the vocabulary is the product, the host is a detail. ## 15 tools **Compose** — `list_catalogs`, `get_catalog`, `required_catalogs`, `distill_document` **Render** — `render_surface`, `render_ping`, `preview_url`, `make_surface_url`, `build_multi_page_surface` **Runbooks** — `emit_runbook_surface`, `emit_training_runbook` (pre-authored compositions; you supply content only) **Ship** — `emit_deployment`, `publish_url`, `unpublish_url` **Setup** — `identify_model` (capability-matched guidance; optional, never a precondition) ## Without MCP at all The vocabulary stands alone. No client, no connector, no account: ```bash curl -X POST https://a2uicatalog.ai/api/render \ -H 'Content-Type: application/json' \ -d '{"blocks":[{"type":"stat_card","value":"1,234","label":"Daily users"}]}' npx -p @a2uicatalog/mcp a2ui render page.json --surface email ``` `OpenAPI` · `spec.json` · `llms.txt` · ARD catalog — all public, all unauthenticated. ## Free No paid tier, because there is nothing to upgrade to. MIT licensed, renderers in the public repo — anything you hit a limit on here, you can run yourself. Full limits: https://a2uicatalog.ai/pricing.md --- [Docs](https://a2uicatalog.ai/llms.txt) · [OpenAPI](https://a2uicatalog.ai/openapi.json) · [Source](https://github.com/a2uicatalog/a2ui) · MIT

First seen 2 Oct 2026. Evidence as of 2 Oct 2026.