Design and UX for AI
Bring your own experience.
AI agents now assemble interfaces at runtime. glph is design and UX for that new surface: you design how software meets you, the brand designs what it may render, and a deterministic engine enforces both, the same way, every run.
The problem
The newest design surface has no designers in it.
You ask an AI agent to compare three plans. It does not answer in prose. It renders an interface: a card, a table, a button, assembled at runtime, shown once, then gone. Nobody designed that screen. The agent picked components from a catalog and filled them with data.
The catalog is the design system of agent UI. It decides which components exist, which variants are legal, which properties are free and which are fixed. The design authority did not disappear. It moved into a file format designers do not touch.
A2UI
Google's open protocol for agents that render UI at runtime from declared component catalogs.
MCP Apps
The first official extension to the Model Context Protocol. Agent-rendered interfaces inside Claude, ChatGPT, and VS Code.
Catalogs
Authored as developer JSON today, with no tooling for designers. The design decisions get made anyway, by whoever writes the file.
How glph works
AI proposes. A deterministic engine disposes.
A model runs only where the problem is genuinely fuzzy: drafting a catalog from your design system, explaining a violation to a human. Deterministic code runs everywhere a guarantee is claimed. Same input, same verdict, every run. The engine never calls a model.
glph Catalog
LiveThe brand side
A designer authoring and governance layer over the component catalogs agents render from. Tokens, scales, contrast floors and an allowlist, checked on every render, with an auto-fix for each violation and a corrected tree you can ship instead.
Open the studioBYOE
LiveThe person side
Bring your own experience. A portable profile of your type floor, contrast floor, line height, density and motion limits, honored everywhere you interface. Same engine, pointed the other way: it runs after the brand pass and has the last word on presentation. Pick a profile in the studio and watch the same surface get re-judged on your terms.
Try a profileThe principles
Five principles of Atomic AI Design.
01
Constraints over generation.
The AI executes within rules. It does not guess.
02
Deterministic over probabilistic.
Same input, same output. Every time.
03
Systems over screens.
Generate the rules, not the pages. A system is infrastructure.
04
Tokens are the API.
Design decisions are code. Not documentation.
05
Open vocabulary, inspectable engine.
The standard is shared. The guarantee is checkable.
Who this is for
Built for the people who own the experience.
Designers who own a design system
Your tokens and rules become a catalog agents cannot render around.
Author it once. The engine holds the line on every surface.
People building agent UI
A deterministic check on every A2UI or MCP App surface, as one API call.
POST a stream, get the verdict, the violations, and a corrected tree.
Anyone who needs their preferences honored
Type floors, contrast floors, reduced motion, carried everywhere you interface.
A preference honored probabilistically is a preference denied.
Stay close
A working proof, built in the open.
glph is one designer directing an AI that proposes, with an engine that disposes. Every guarantee lands in deterministic TypeScript pinned by a public test suite, the engine answers at POST /api/check, and the essay makes the argument. If you work on agent UI, governance, or the protocols underneath them, I want to compare notes. Find me at michaelhiggins.ai.