Your AI keeps inventing buttons. Here's how to stop it.

Ask an AI coding assistant for a settings page and you will get one — with a brand-new Button, a bespoke Card, and a hand-rolled table that almost matches the three you already have. Multiply by every prompt across every engineer, and your codebase becomes a museum of almost-identical components.

The model isn't wrong. It's blind.

LLMs generate from training priors unless you give them something better. Your component library is invisible to the model: it lives in your repo, your Storybook, your Figma file — none of which the model consults mid-generation. Hallucinated components are not a model failure; they are a context failure.

Give the model eyes: a component catalog over MCP

The Model Context Protocol lets a local server expose tools any MCP-capable client — Claude Code, Cursor, and others — can call. Pegasus ships one that answers three questions: what components exist (search_components), how each works (get_component: props, variants, tokens, usage), and how they compose into common pages (get_block).

With the catalog in hand, generation changes shape. The model queries before it writes. A settings page becomes Tabs + Form + DataTable from your actual library, themed by your actual tokens. Zero invented components, nothing to repaint afterwards.

Why local matters

The server runs on your machine, next to your checkout. No code leaves your environment, the catalog is always in sync with the version you installed, and it works in a coffee shop with bad wifi. It is infrastructure, not a SaaS dependency.

The compounding effect

Every AI-generated page that reuses real components makes the next one cheaper to review, because reviewers check composition, not implementation. Design systems were always leverage. Wired into your AI tools, they become compounding leverage.