Scope
This is Velofy’s open source software (OSS) map from the September 23, 2026 research survey. Star counts are GitHub counts as of September 2026. Licenses and roles are recorded as found in that survey. “None (no LICENSE file)” and NOASSERTION are not permissive licenses: no open source license was found or identified, so reuse rights cannot be assumed.
First principles
An artificial intelligence (AI) agent that builds a user interface (UI) needs four things: values (colors, type, spacing), guidance on when to use them, components to build with, and a way to install or fetch all three. The sections below map the layers that supply them, plus the evidence those layers are built from.
- Values are data. Data travels best in a typed, shared format, which is why the token layer has converged on the Design Tokens Community Group (DTCG) format.
- Guidance is prose. It lives in
DESIGN.mdandSKILL.mdfiles. Prose has no type system, so formats in these layers drift and multiply. - Components and motion are code. Code can be run, tested, and versioned, and these layers are older and more settled than the
DESIGN.mdlayer. - Evidence about real products is someone else’s work. Screenshots and extracted styles come from products the collector did not make, so licensing limits what can be shared. That layer is the thinnest.
Two columns in every table need care. Stars measure attention at one moment, not correctness, and a few accounts can dominate them. License decides whether a project can be reused at all: a public repository without a license grants no reuse rights by default.
1. Stitch DESIGN.md format
The format layer is small. Google Labs supplies the reference specification; the other projects show local-first use and independent iteration.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| google-labs-code/design.md | Apache-2.0 | 28,047 | Google Stitch and Google Labs’ source-of-truth visual-identity format for coding agents. |
| soulcore-dev/soul-design-md | MIT | 2 | A v2.0 fork adding component states, motion specifications, accessibility notes, and inline snippets. |
| ricocc/ricoui-design-md | Apache-2.0 | 97 | A local-first design-system workspace that treats DESIGN.md as a read/write source of truth. |
2. Awesome lists and design-system references
These are discovery and reference layers. They range from reverse-engineered DESIGN.md collections to long-running lists of real-world systems and token tooling.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| VoltAgent/awesome-design-md | MIT | 117,378 | Curated reverse-engineered brand DESIGN.md collection. |
| bergside/awesome-design-skills | MIT | 2,894 | Cross-tool index of 67 DESIGN.md and SKILL.md design skills. |
| VoltAgent/official-design-md | MIT | 508 | Collection restricted to first-party company-published DESIGN.md files. |
| alexpate/awesome-design-systems | Unlicense | 25,997 | Long-running list of real-world design systems, component libraries, docs, and pattern libraries. |
| sturobson/Awesome-Design-Tokens | Unlicense | 1,299 | Curated material on design-token tools, naming, and standards. |
3. shadcn/ui and registry tooling
The registry is an installation protocol as much as a component catalog. These projects show the core system, an implementation example, a directory, and a teaching registry.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| shadcn-ui/ui | MIT | 124,444 | Copy-paste component source, command-line interface (CLI), and JavaScript Object Notation (JSON) registry format for installable components and themes. |
| ncdai/chanhdai.com | MIT | 2,289 | Personal portfolio and working custom shadcn registry reference implementation. |
| rbadillap/registry.directory | MIT | 76 | Directory and explorer for the shadcn-registry ecosystem. |
| WebDevSimplified/wds-shadcn-registry | MIT | 292 | Educator-maintained registry used to teach registry authoring. |
4. Design-token specifications and tooling
The World Wide Web Consortium (W3C) DTCG format is the converged token layer. The projects here write, validate, transform, or generate from the $type and $value token format.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| design-tokens/community-group | NOASSERTION (W3C document terms, no OSS LICENSE file) | 2,127 | W3C Design Tokens Community Group repository for the Design Tokens Format Module. |
| style-dictionary/style-dictionary | Apache-2.0 | 4,823 | Build system that transforms tokens into platform-specific output and supports DTCG. |
| tokens-studio/figma-plugin | MIT | 1,608 | Tokens Studio for Figma, a common authoring and sync path for tokens. |
| terrazzoapp/terrazzo | MIT | 459 | DTCG-first generator for web, mobile, and native code. |
| sugarcube-sh/sugarcube | MIT | 127 | Smaller tool that turns design tokens into Cascading Style Sheets (CSS) and components. |
| nclsndr/design-tokens-format-module | MIT | 41 | Independent TypeScript implementation and validator of the DTCG format. |
5. Design skill packs for AI agents
This category has broad adoption but no design-native standard. Most projects pour design process and taste guidance into a general SKILL.md container.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| anthropics/skills | None (no LICENSE file) | 177,692 | Anthropic’s public Agent Skills repository and the foundational packaging convention. |
| obra/superpowers | MIT | 290,299 | Agentic-skills framework and software-development methodology. |
| Nutlope/hallmark | MIT | 29,049 | Anti-AI-slop design skill for Claude Code, Cursor, and Codex. |
| Owl-Listener/designer-skills | MIT | 2,724 | Skills, slash commands, and plugins for design research, systems, UI, interaction, and delivery. |
| julianoczkowski/designer-skills | Apache-2.0 | 560 | Process-oriented skills for designers using AI coding tools. |
| elayadesign/ai-design-skills | MIT | 2,302 | General-purpose AI design-skills collection with one landing-page skill at the research date. |
6. Motion and transition libraries
Motion tooling is more mature than the DESIGN.md ecosystem. The layer includes broad animation libraries, a zero-config helper, and browser View Transitions support.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| motiondivision/motion | MIT | 33,697 | Framework-agnostic successor to Framer Motion for React and vanilla JavaScript. |
| greensock/GSAP | None (custom license terms) | 28,574 | Timeline and scroll-animation platform for complex sequenced motion. |
| formkit/auto-animate | MIT | 13,922 | Zero-config utility for smooth animation of Document Object Model (DOM) changes across frameworks. |
| shuding/next-view-transitions | MIT | 2,387 | Next.js App Router wrapper for the native View Transitions application programming interface (API). |
| demarketed/view-transitions-polyfill | ISC | 72 | View Transitions API polyfill for engines without native support. |
7. Component kits and starters
These are the implementation layer beneath many generated interfaces: accessible primitives, typed styling systems, and source-ready component collections.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| radix-ui/primitives | MIT | 19,315 | Unstyled accessible React primitives and a foundation for shadcn/ui. |
| mui/base-ui | MIT | 10,973 | Unstyled UI components from the Radix, Floating UI, and Material UI teams. |
| chakra-ui/panda | MIT | 6,195 | Type-safe CSS-in-JavaScript (CSS-in-JS) framework built for design systems. |
| vanilla-extract-css/vanilla-extract | MIT | 10,429 | Zero-runtime CSS-in-TypeScript compiled to plain CSS. |
| untitleduico/react | MIT | 1,939 | Tailwind and React Aria component collection matched to a paid Figma system. |
| origin-space/ui-experiments | None (no LICENSE file) | 1,427 | Full-page layouts and UI experiments built with Origin UI and shadcn/ui. |
8. Open UI pattern and screenshot datasets
This is the thinnest category. The available work is a live Model Context Protocol (MCP) scraper, a curated gallery, and a small stale desktop dataset rather than a durable open corpus of real product UI.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| INSANE0777/Awwards-mcp | MIT | 47 | MCP server for live award-site inspiration, design-DNA extraction, and page capture. |
| inbn/component-gallery | None (no LICENSE file) | 348 | Hand-maintained reference collection of components from established design systems. |
| waltteri/desktop-ui-dataset | None (no LICENSE file) | 8 | Small annotated desktop screenshot dataset in PASCAL Visual Object Classes (VOC) Extensible Markup Language (XML), last pushed in 2020. |
9. Style and design extractors
Extraction is fragmented. The projects cover Chrome, Figma, running products, and iOS apps, with no shared accuracy benchmark or dominant pipeline.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| bergside/design-md-chrome | MIT | 2,886 | Chrome extension that extracts computed styles and creates DESIGN.md and skill files. |
| bergside/design-md-figma | MIT | 151 | Figma sibling that creates DESIGN.md and SKILL.md from local style guidelines. |
| conorluddy/tokextract | None (no LICENSE file) | 3 | Extracts DTCG tokens, a DESIGN.md narrative, and an audit from vibe-coded iOS apps. |
| 20prateeksingh/design-context-for-ai | MIT | 7 | Maps navigation, captures editable snapshots, and extracts product tokens for AI context. |
10. Literal DESIGN.md convention and design-engineering repositories
These projects use the vocabulary directly. They are early, low-adoption examples of tooling and agents framed around design engineering itself.
| Name | License | Stars as of September 2026 | Role |
|---|---|---|---|
| samuelbifalco/SkillMaker | MIT | 0 | Builder for SKILL.md and DESIGN.md agent documents. |
| ddruids/design-engineer | MIT | 10 | Claude skill combining Figma-to-frontend work with a React, Next.js, and TypeScript handbook. |
| AgentsORG/Benji | None (no LICENSE file) | 2 | Early design-engineering agent with persistent design memory and a checkable constitution. |
What the map says
The token format layer has a clear center in DTCG. The DESIGN.md layer is young and highly visible, but activity is concentrated in a few accounts. Skills are common packaging, not a settled design-specific format. Extraction and open UI datasets remain fragmented and thin.
For any future Velofy work, this map supports a narrow proposal: extend existing formats, label provenance and confidence, and publish direct artifacts that both people and agents can inspect. It does not describe a shipped Velofy design package, catalog, or MCP server.