Meta Open-Sources Astryx: Design System Built for AI Agent Integration

Eight Years of Internal Use

Astryx is not a new project created for open source. Meta used it internally for eight years to support over 13,000 applications and more than 150 components before releasing the public parts. The components are not demo code; buttons, forms, dark mode, and accessibility features have all run in real production workloads, with accessibility support driven by actual internal requirements.

Unified Interface for Humans and Agents

Previously, humans, build tools, and AI agents consumed three different interfaces: humans read documentation sites, build tools parsed configuration files, and agents had no dedicated entry point and had to guess usage from web pages. Astryx solves this by placing component documentation, design tokens, page templates, theme utilities, and upgrade codemods into a single CLI. The official documentation states that the API called by AI agents and build tools is the same as the CLI.

Installation runs npx @astryxdesign/cli init, which also generates AGENTS.md or CLAUDE.md. The command has no interactive prompts, so it can be executed directly by an agent.

How Styling Works

Astryx uses StyleX for styling. Developers create style objects with stylex.create and pass them to components via the xstyle prop, eliminating className concatenation.

// Old way: className string concatenation, styles in global CSS, overrides require priority management
<Button className="btn btn--primary">Submit</Button>

// New way: stylex.create generates style objects, xstyle receives the object
const styles = stylex.create({
  primary: { paddingBlock: 8, paddingInline: 16 } // compiled to atomic CSS at build time
});
<Button xstyle={styles.primary}>Submit</Button>

Key Insight

The real problem Astryx solves is not a lack of good-looking components. It encodes component usage as machine-readable interfaces so that humans and agents receive the same specifications. For teams already using AI agents to build UIs, making documentation, tokens, and command-line tools machine-readable is more valuable than the components themselves.