Skip to content

Now in the official shadcn registry directory.

Start free
ai2v0.7
Perspective

Why we built an agent-native design system

Most registries are written for humans and merely tolerated by agents. We flipped that: the metadata is written to be read by models, and it is the same metadata people read.

4 min read
Share

An AI agent installing UI is no longer a demo. It is how a lot of code gets written now. So the question we kept asking was simple: when an agent reaches for a component, what does it actually see?

Code-only is not enough

A registry that serves source files answers half the question. The agent gets the code, but not the intent: which variant to use, what a tone means, whether a size matches the button next to it. It guesses, and guessing is where the tokens get burned.

The cheapest token is the one an agent does not have to spend figuring out what you meant.

What agent-native means here

Every item states its variant matrix, intended use and dependencies in plain words. That description is not a second artifact for machines. It is the same text a person reads on the component page.

  • A working example per component, so the agent copies real usage instead of inventing props.
  • Dependencies, including the token theme, resolve automatically through registryDependencies.
  • The whole surface is indexed in an llms.txt file for tools that do not speak MCP.

It is the same system either way

The result is one 51-component core that a person browses and an agent installs, from the same registry, with the same metadata. No separate machine mode. Just a system that was honest about who reads it.

Back to the blog