UI for Agents

Half the code written against Mizu will be written by an agent. The system ships three resources so Cursor, Claude Code, Codex, and friends use it correctly without you explaining anything.

1. Drop AGENTS.md into your project

AGENTS.md is a README for agents: most coding agents read it from your project root automatically. Mizu maintains a canonical one at https://mizu-ui.com/AGENTS.md covering install commands, the composition patterns, and the design rules an agent must not violate. One command adds it:

It is generated from the same catalog as the docs, so it always matches the library. Re-run the command after a Mizu update to refresh it.

2. Point tools that use their own file

Claude Code reads CLAUDE.md first, and Cursor reads .cursor/rules. A one-line pointer is enough; both follow it:

# CLAUDE.md
See AGENTS.md for the Mizu design system rules and install commands.

3. The library is machine-readable

https://mizu-ui.com/llms.txt lists every component and block with its description, docs link, and registry URL. The registry itself serves clean JSON per item, so an agent can discover a component and install it with its whole dependency tree in one command. If you are prompting an agent cold, this is all it needs:

Read https://mizu-ui.com/llms.txt for the component catalog,
then install what you need with:
npx shadcn-svelte@latest add https://mizu-ui.com/r/v0.1.4/<slug>.json

What the agent gets right as a result

• Installs through the registry instead of pasting stale component code from its training data.
• Composes assistant screens in the right order, following the Build a chat anatomy, or installs the finished block.
• Respects the design rules: no decorative strokes, one accent, auras only for AI moments, nothing heavier than semibold, motion behind prefers-reduced-motion.
• Themes through tokens instead of scattering hex values through your markup.