**The vocabulary layer for A2UI — useful for humans · declarative for agents.** The model names an atom. The renderer compiles the HTML, CSS, SVG, and animation. [![Atoms](https://img.shields.io/badge/atoms-474-7c9cff?style=flat-square&labelColor=0a0e17)](https://a2uicatalog.ai/) [![Surfaces](https://img.shields.io/badge/surfaces-8-38bdf8?style=flat-square&labelColor=0a0e17)](https://a2uicatalog.ai/spec.json) [![MCP](https://img.shields.io/badge/MCP_server-a2uicatalog.ai%2Fmcp-7c9cff?style=flat-square&labelColor=0a0e17)](https://a2uicatalog.ai/mcp) [![ARD](https://img.shields.io/badge/ARD-ai--catalog.json-38bdf8?style=flat-square&labelColor=0a0e17)](https://a2uicatalog.ai/.well-known/ai-catalog.json) [![License](https://img.shields.io/badge/license-MIT-34d399?style=flat-square&labelColor=0a0e17)](LICENSE) [![A2UI](https://img.shields.io/badge/spec-v1.0_candidate-a78bfa?style=flat-square&labelColor=0a0e17)](renderers/a2ui_v1.py) *Independent, unofficial catalog — not affiliated with or endorsed by Google. A2UI is Google's protocol; the official spec lives at [a2ui.org](https://a2ui.org).*

--- ## The idea Rather than asking an agent to generate custom UI every turn — expensive, fragile, unpredictable — give it a stable vocabulary of atoms and let it compose from those. ``` Raw HTML 609 tok ████████████████████████████████████████ OpenUI 287 tok ███████████████████ A2UI 68 tok ████ ``` Fewer tokens to describe the same UI. The model names an atom; the renderer expands it into the full HTML server-side, and that expansion never re-enters the model's context window. (The efficiency framing here is still being validated — see `benchmarks/BENCHMARK.md` for the current methodology.) > **Work in progress.** The atom vocabulary itself is stable; what it's *for* is still being explored. The live catalog and this repo are the vocabulary and its renderers — applying that vocabulary to new use cases (rendering a CLI's own output, for instance, in a sibling project not yet published) is ongoing, not finished. --- ## Connect an agent The catalog is consumable three ways, no signup for any of them — BYO by design: renders happen in *your* environment, your data never touches us. | Channel | Endpoint | What an agent gets | |---|---|---| | **MCP server** | [`a2uicatalog.ai/mcp`](https://a2uicatalog.ai/mcp) | Tools to browse catalogs, validate/compose payloads, emit render URLs, and render live views on MCP Apps hosts | | **ARD manifest** | [`/.well-known/ai-catalog.json`](https://a2uicatalog.ai/.well-known/ai-catalog.json) | Agentic Resource Discovery — every atom as a crawlable, typed entry | | **Machine catalog** | [`a2uicatalog.ai/spec.json`](https://a2uicatalog.ai/spec.json) | The full schema in one fetch — types, fields, surface compatibility, degradation notes | **5-minute quickstarts:** [Claude (Desktop / claude.ai)](docs/quickstart-claude.md) · [Gemini Enterprise (BYO-MCP)](docs/quickstart-gemini-enterprise.md) **As an agent skill** — this repo ships a [`SKILL.md`](SKILL.md), so any agent using the [skills](https://skills.sh) ecosystem can pick up the catalog's when-to-use guidance directly: ```bash npx skills add a2uicatalog/a2ui@a2ui-atomic-catalog ``` A few atoms, live on the site — click through to a full spec + fields table for each:

glowing_stat

Command Palette

Tagged “generative-ui”

Servers

a2ui