Browse projects from 2010 to today. Use tabs to filter and keyboard shortcuts to navigate quickly: ↑/↓ selects, PgUp/PgDn changes page, A-Z jumps by letter.
ITEM: 1/27
Cactus Design System MCP
Company...: Cactus Gaming
Period....: Aug 2026 — Present
Category..: FULLSTACK
Stack:
- TypeScript
- MCP
- Next.js
- Vercel
- OAuth 2.1
- PKCE
- GitHub
- Zod
- Design Systems
- Developer Experience
Summary:
Grounding AI agents on a private design system: a lookup service for Cursor, VS Code and Claude so models use the real UI kit instead of inventing it. Models never learned the 58 components, so they guessed sizes, properties and import paths.
Highlights:
- On every kit build we publish a catalog: each component's real properties, color/spacing tokens, the correct import, usage and accessibility notes when they exist, and the changelog
- If a note is missing, the AI admits the gap — it does not invent a spec. The build fails if any of the 58 components is missing its property list
- On the laptop, the AI matches that product's kit version. In the cloud, people sign in with GitHub; only our org gets access, so designers can use it with no install
- Vs. the first plan: 13 AI queries → 6; cloud + company login in v1 (was later); real catalog from day one, not a demo dataset
- A new kit release opens a PR to update this service and re-check the catalog
- Does not replace Figma, Storybook or review. Stops the AI from lying about our components