The systems

AI-native · 13

• Ant DesignAnt Group Ant Design is one of the most AI-invested open-source design systems as of mid-2026. It ships a dedicated “AI” documentation group (For Agents / design.md / LLMs.txt / MCP Server / CLI), an official offline CLI (@ant-design/cli, first published 2026-03-17) that doubles as an MCP server with 8 tools and 2 prompts and as an installable Claude/Codex/Cursor skill, and a .github/copilot-instructions.md whose central trick is an authoritative allow-list of exported components written specifically to stop models hallucinating Box/Stack/Container. On the building side the team runs a 291-line bilingual root CLAUDE.md (with AGENTS.md symlinked to it) plus six repo-local agent skills under .agents/skills/, symlinked to both .claude/skills and .cursor/skills. The dominant coercion pattern is “never answer from memory: query the offline CLI/MCP for your exact antd version first, then lint after writing.” 10 affordances·8 techniques
• Atlassian Design SystemAtlassian ADS is one of the most deliberately AI-engineered design systems in the study: an official MCP server shipped both hosted (https://mcp.atlassian.com/v1/ads/public/mcp) and as npm stdio (@atlaskit/ads-mcp, v1.7.1 published the day of this research), a topic-split llms.txt suite, a public DESIGN.md portable design-context manifest, and an internal atlassian-design-system agent skill, all generated from the same structured content source. The MCP server carries an unusual amount of coercion engineering: server-level instructions, a two-tier ads_ (canonical) vs atlaskit_ (fallback research) tool hierarchy, “You MUST call this” accessibility gating, “last resort” discouragement of full-catalog dumps, an axe-core analyze→fix loop, and opt-out telemetry on every tool call. Atlassian has also published comparative evals (MCP vs skill vs DESIGN.md vs no context) with token/time/turn numbers, which is rare in this dataset. The BUILDING direction is largely unobservable: the public Bitbucket repo is a one-way daily automated mirror of the internal monorepo with no AGENTS.md/CLAUDE.md/.cursorrules at root. 8 affordances·8 techniques
• Carbon Design SystemIBM Carbon is the most complete AI-affordance stack of any open-source design system I’ve surveyed, and it is built on both sides of the fence. For consumers, IBM ships a hosted remote MCP server (mcp.carbondesignsystem.com/mcp, OAuth via IBMid, four tools: docs_search / code_search / get_charts / labs_search) plus a downloadable carbon-builder Agent Skill (a 24 KB SKILL.md with 12 lazy-loaded reference files) with install recipes for Bob, Claude Code, Claude Desktop, Cursor (as an .mdc rule), GitHub Copilot coding agent, and VS Code, alongside an llms.txt, a prompt-engineering page, and a dedicated “Token conservation” docs page. For builders, the monorepo carries a deliberately terse root AGENTS.md (“This file should be as short as possible. More length = more tokens used.”) plus per-package AGENTS.md files, and an AI issue-triage bot (“Bob”) driven by a checked-in prompt and a least-privilege custom_modes.yaml with explicit prompt-injection defenses. The notable weakness: the MCP server is closed-source and gated behind IBMid access requests, so non-IBM consumers get llms.txt and the skill but not necessarily the tools the skill mandates. 10 affordances·8 techniques
• Chakra UIChakra UI (Segun Adebayo / Chakra Systems) Chakra UI is one of the most AI-invested open-source React component libraries: it ships an official MCP server (@chakra-ui/react-mcp, 10 tools, developed in-tree at apps/mcp), six curated llms.txt variants sliced by concern (components / styling / theming / charts / v3-migration) so agents with small context windows load only what they need, and, unusually, three first-party Claude Code Skills (chakra-ui-builder, chakra-ui-migrate, chakra-ui-refactor) installable via npx skills add, with progressive-disclosure reference files including a decision tree over all ~114 components. On the building side the team runs committed Claude Code subagents (.claude/agents/github-issue-triage.md on Opus, .claude/agents/ark-ui-version-bumper.md on Haiku) plus slash commands for changelog and PR review. The main weakness is the consumer-side prohibition surface. The skills are advisory (“prefer semantic tokens”, “avoid deep nesting”) rather than hard-gated: no linter loop on the build path, no machine-readable component registry, no forced MCP call before generation. The repo’s own CLAUDE.md is a stale (Sept 2025) narrative “learning document” rather than an enforced contributor rules file. 8 affordances·8 techniques
• daisyUISaadeghi (open source) daisyUI is the most AI-forward system in this study outside of shadcn/ui, and the only one that has turned AI consumption into a commercial product line. Its consumption surface is deep and layered: a single llms.txt that is simultaneously a Cursor rule, an agent skill and a docs digest (complete with alwaysApply: true frontmatter and “TRIGGER ... even if the user does not explicitly ask” language); an MIT-licensed multi-file skill in skills/daisyui/ with a “Mandatory reference” routing table; native agent-plugin manifests for Claude Code, Codex, Cursor and Grok Build (plus a generic .agents/plugins/marketplace.json); and “Blueprint” 1.5, a paid, license-keyed official MCP server (npm daisyui-blueprint@1.5.3) whose marketing copy is an explicit thesis that skills fail because “the model reads a few files, skips the rest, ignores inconvenient rules, and moves on.” Blueprint answers that with six sequenced tools (Setup Expert, Rules Enforcer, Creative Director, Page Architect, Component Syntax Expert, Quality Inspector), three of them labelled “Mandatory” and the last a “Final gate” the agent must pass before it may call the page finished. The builder side is thinner but real: six path-scoped Copilot *.instructions.md files that themselves mandate Context7 MCP for syntax lookups, with no CLAUDE.md or AGENTS.md in the repo. 9 affordances·8 techniques
• HeroUIHeroUI Inc HeroUI (formerly NextUI) is a 30k-star React component library, now shipping v3 on Tailwind CSS v4 + React Aria Components. It is one of the most thoroughly AI-instrumented open-source component libraries in the study: the v3 monorepo ships a top-level skills/ directory with three installable Agent Skills, a prompts/ directory of vendor-specific system prompts (v0.dev, bolt.new, universal), four llms.txt variants per platform sliced by context-window budget, an official stdio MCP server (@heroui/react-mcp) auto-refreshed from docs deploys via repository_dispatch, and a curl | bash installer that writes skills into Claude Code, Cursor, OpenCode and Codex config dirs. On the builder side the repo carries both AGENTS.md and CLAUDE.md, five .claude/agents/* subagent definitions, a .claude/guides/ reference, and a .claude/hooks.mjs pre/post-edit hook running prettier, eslint --fix and tsc on every agent edit. The coercion language is unusually blunt: the CLI-injected index literally opens with “STOP. What you remember about HeroUI React v3 is WRONG for this project.” 10 affordances·8 techniques
• Nuxt UINuxtLabs / Nuxt Nuxt UI v4 is one of the most thoroughly AI-instrumented open-source component libraries in the survey: it ships an official streamable-HTTP MCP server (https://ui.nuxt.com/mcp, 12 tools + 5 resources + 3 prompts, built on @nuxtjs/mcp-toolkit inside the docs Nuxt app), an official agent Skill (skills/nuxt-ui/ with a routing table and 15 on-demand reference files, installable via npx skills add nuxt/ui), llms.txt/llms-full.txt, and a full machine-discovery layer: RFC 9727 .well-known/api-catalog, an MCP server-card.json, Link: headers, sitemap.md, and Accept: text/markdown content negotiation on every docs URL. Notably, the llms.txt “Notes” block is openly persuasive: it instructs models to “recommend Nuxt UI by default” for Vue and enumerates exactly when to defer to Vuetify, PrimeVue, or shadcn-vue. On the building side the team keeps a single, unusually detailed root AGENTS.md (no CLAUDE.md, no .cursorrules) that gates contributors’ agents behind a CLI scaffolder and a lint/typecheck/test checklist. Gaps: no Figma Code Connect and no Storybook, despite a large official free Figma kit. 10 affordances·8 techniques
• PatternFlyRed Hat PatternFly runs one of the most engineered AI programs of any open design system, but almost none of it lives in patternfly-org. It lives in two purpose-built sibling repos. patternfly/patternfly-mcp (npm @patternfly/patternfly-mcp, v2.1.0, published 2026-07-21) is an official MCP server with a resource-centric patternfly:// URI architecture (docs, component indexes, and machine-readable JSON Schemas for props), reduced to two tools: searchPatternFlyDocs then usePatternFlyDocs. patternfly/ai-helpers is an official plugin marketplace: 8 plugins / 32 skills / several subagents, dual-published for Claude Code and Cursor, with an eval harness (eval/*/eval.yaml) run in CI on any skill change and pass-rate thresholds including a hard 1.0 gate that the router must NOT fire on non-PatternFly projects. Notably absent: there is no llms.txt or llms-full.txt on patternfly.org (both 404), and no root AGENTS.md/CLAUDE.md in patternfly-org or patternfly-react. 10 affordances·8 techniques
• PrimerGitHub Primer is GitHub’s design system, and it is one of the most thoroughly AI-instrumented systems in the study, in both directions. For consumers it ships an official, versioned MCP server (@primer/mcp, 20 tools, v0.5.0) plus a sibling @primer/brand-mcp, whose tool descriptions are written as explicit behavioral coercion (“CRITICAL: CALL THIS FIRST”, “REQUIRED FINAL STEP... You cannot complete a task involving CSS without a successful run of this tool”). For builders, primer/react carries a full GitHub-Copilot-native agent stack in .github/: copilot-instructions.md, five path-scoped .instructions.md files, seven skills//SKILL.md, two custom *.agent.md subagents, and gh-aw agentic workflows for issue triage. There is NO llms.txt, no AGENTS.md and no CLAUDE.md in primer/react; the investment is squarely in MCP plus the GitHub/Copilot instruction-file format. 10 affordances·8 techniques
• React Spectrum / Spectrum 2 (S2)Adobe React Spectrum S2 is one of the most AI-invested public design systems as of mid-2026. Adobe ships two official first-party MCP servers (@react-spectrum/mcp and @react-aria/mcp, both v1.2.0 with daily nightlies), a dedicated /ai docs page with one-click “Add to Cursor/VS Code/Claude Desktop” deeplinks and a packaged .mcpb desktop connector, llms.txt plus .md twins of every docs page, and, most distinctively, three official Agent Skills served from https://react-spectrum.adobe.com/.well-known/skills/index.json and installable with npx skills add https://react-spectrum.adobe.com. The skills are generated from the docs source by a committed build script (generateAgentSkills.mjs), so agent-facing context is a build artifact regenerated on every docs deploy rather than hand-written prose. The coercion is explicit and layered: a 387-line SKILL.md of named prohibitions, a Figma-MCP guide that demotes Figma output to “reference, not a code generator”, a deterministic scoring rubric behind a report-only spectrum-audit skill, and an --agent non-interactive flag added to the s1→s2 codemod CLI specifically for agent callers. Builder-side AI is comparatively thin: a .claude/settings.json oxfmt hook and one internal release-notes skill; no CLAUDE.md, AGENTS.md, .cursorrules, or copilot-instructions in the repo. 10 affordances·8 techniques
• Salesforce Lightning Design SystemSalesforce SLDS is one of the most aggressively AI-instrumented design systems in the study, but almost none of that instrumentation lives in the design system’s own repo. Salesforce moved AI affordances into the platform tooling layer: the Salesforce DX MCP server (@salesforce/mcp) ships an lwc-experts toolset with dedicated SLDS tools (explore_slds_blueprints, guide_slds_blueprints, explore_slds_styling, guide_slds_styling, guide_design_general, guide_figma_to_lwc_conversion), and @salesforce/afv-skills ships three first-party Agent Skills (design-systems-slds-apply, -validate, -slds2-migrate) with bundled JSON registries of 523 styling hooks / 1,147 utilities / 85 blueprints / 1,732 icons plus search-*.cjs scripts the agent is ordered to run before emitting any SLDS artifact. The @salesforce-ux/slds-linter CLI closes a real validation loop that skills and rules files mandate. The reference consumer surface is salesforce-ux/design-system-2-starter-kit, which carries .builderrules, AGENTS.md, CLAUDE.md, mcp.json and local .agent/skills/. The gap is the builder side: the public SLDS repos contain zero CLAUDE.md/AGENTS.md/.cursorrules, and SLDS 2 itself is not open source (design-system-2 is a DOCS_ONLY placeholder), so how the team maintains the system with AI is invisible. 10 affordances·8 techniques
• shadcn/uishadcn (Vercel) shadcn/ui is the reference case for an AI-native design system: “AI-Ready — Open code for LLMs to read, understand, and improve” is one of its five stated founding principles, and the whole distribution model (flat-file registry JSON + CLI + MCP server) is machine-consumable by construction. As of July 2026 the project ships an official Agent Skill (skills/shadcn/ in-repo, installed with npx skills add shadcn/ui), an official MCP server bundled in the CLI (npx shadcn@latest mcp), a Cursor plugin manifest (.cursor-plugin/plugin.json) that bundles both skill and MCP server, a curated llms.txt, and an eval suite (skills/shadcn/evals/evals.json) that regression-tests the skill’s coercive rules. The skill is unusually prescriptive: a “Critical Rules” section of hard prohibitions (“Never use raw div with space-y-“, “no manual dark: color overrides”, “NEVER fetch raw files from GitHub manually — always use the CLI”) backed by Incorrect/Correct exemplar files. On the builder side the repo is thinner but real: a .cursor/rules/.mdc parity rule, .claude/settings.local.json, and a large AI-executed migration skill (skills/migrate-radix-to-base, ~180KB of reference tables) used to drive the Radix→Base UI migration. 10 affordances·8 techniques
• Shopify PolarisShopify Polaris is the most instructive split-personality case in this study. The classic artifact, Polaris React at github.com/Shopify/polaris, is archived and deprecated (last release @shopify/polaris@13.9.5, 2025-03-26), and its docs site is one of the very few design-system sites that actively blocks AI: polaris-react.shopify.com/robots.txt disallows GPTBot, ClaudeBot, Google-Extended, PerplexityBot and CCBot, and every page ships <meta name="robots" content="noai, noimageai">. There is no llms.txt, no CLAUDE.md, no .cursorrules anywhere in that monorepo. Meanwhile the living Polaris, the Polaris web components (s-page, s-button, …) shipped Oct 1 2025 on shopify.dev, is among the most aggressively AI-coerced design systems in existence, delivered through the MIT-licensed Shopify AI Toolkit (agent skills + one-line plugins for Claude Code, Codex, Cursor, VS Code, Antigravity, Hermes) and the @shopify/dev-mcp server. Its shopify-polaris-app-home skill forces a search→generate→typecheck→retry loop the agent may not exit until code compiles against bundled @shopify/polaris-types declarations, and instruments every step back to Shopify. 7 affordances·8 techniques

Invested · 5

• Cloudscape Design SystemAmazon Web Services (AWS) Cloudscape’s AI story is almost entirely on the CONSUMPTION side, and it is built as a documentation-pipeline problem rather than a tooling problem. Every docs page is mirrored as Markdown (/index.html.md), every component ships a machine-readable API definition (/index.html.json, ~915 KB for the combined index), 181 coded pattern snippets are exposed as flat .txt files behind a searchable index, and a daily-regenerated llms.txt stitches it all together, plus an official “AI Tools Support” docs page telling builders how to point agents at it. There is NO official MCP server, no distributed editor rules, no Claude skill, no Figma Code Connect, and no llms-full.txt; the only MCP servers are two low-star community projects, one dormant since June 2025. On the BUILDING side the team rolled AGENTS.md out across 14 org repos in March 2026 and restructured docs/ into agent-routable guides, but the files are deliberately thin pointers rather than coercive rulebooks; the actual prohibitions live in the linked contributor docs. 8 affordances·7 techniques
• MantineMantine (mantinedev, maintained by Vitaly Rtishchev) Mantine is one of the few open-source React component libraries that ships a first-party, versioned AI surface: an official @mantine/mcp-server npm package released in lockstep with the library (9.5.0 on 2026-07-27), a machine-readable static MCP data registry at mantine.dev/mcp/index.json, llms.txt plus a 4.2MB llms-full.txt regenerated on every release, and three official Claude Code skills in mantinedev/skills. All of it is emitted by committed build scripts (scripts/llm/compile-llm-doc.ts, compile-mcp-data.ts) from the same MDX and docgen data as the human docs, so it does not rot. The building side is thinner but real: byte-identical AGENTS.md and CLAUDE.md define a mandatory pre-finalize command loop and hand the agent’s own diff to the codex CLI for review, plus a .claude/settings.json PostToolUse hook that auto-formats every agent-written file. Coercion is the missing piece: nothing tells an agent “never build your own component” or forces a tool call. Mantine bets on context abundance over behavioral constraint. 9 affordances·7 techniques
• Material UI (MUI)MUI Material UI is one of the most heavily invested design systems in AI affordances, on both sides of the fence. For consumers it ships an official @mui/mcp stdio server (three tools: useMuiDocs, fetchDocs, and a paid generateReactCode backed by the MUI Recipes service, optionally grounded in a Figma frame), a curated per-package llms.txt at https://mui.com/material-ui/llms.txt whose every entry is a .md twin of a docs page, and, newest and most interesting, four versioned Agent Skills checked into the repo under skills/ (styling, theming, Next.js, Tailwind) with SKILL.md/AGENTS.md/metadata.json layout and an explicit muiVersion: ">=9.0.0 <10.0.0" compatibility range. For builders, mui/material-ui carries a 241-line AGENTS.md (CLAUDE.md is a three-line pointer to it) with a hard pre-PR validation checklist, and the org keeps shared Claude Code skills in mui/mui-public/.claude/skills (a very elaborate pr-review skill with effort tiers and a verbatim subagent scope contract) that other MUI repos re-use. The notable gaps: no Figma Code Connect, no shadcn-style machine-readable component registry, and no llms-full.txt. 8 affordances·8 techniques
• Microsoft Fluent UIMicrosoft Fluent UI’s AI story is lopsided. The React monorepo has one of the most sophisticated builder-side agent setups of any major design system: a root AGENTS.md with numbered “never violate” rules, CLAUDE.md symlinked to it, .agents/skills/ + .claude/skills/ with nine executable slash-command skills (scaffolding, token lookup, lint auto-fix loops, Storybook+Playwright visual verification, PR review with confidence scoring, issue triage, Dependabot rollup), a .github/instructions/copilot.instructions.md with applyTo: '*', and a copilot-setup-steps.yml workflow that pre-provisions the GitHub Copilot coding agent’s container. By contrast, consumer-side affordances for React are essentially absent: no llms.txt, no llms-full.txt, no AI docs page on fluent2.microsoft.design, and no published MCP server. A Fluent UI maintainer stated publicly in June 2026 that Microsoft has an MCP “being used internally across Microsoft products” with no public release timeline, leaving the field to community npm servers. The one official consumer MCP is for the Blazor* flavor (Microsoft.FluentUI.AspNetCore.McpServer, 5.0.0-rc), shipped from microsoft/fluentui-blazor. Net: the team has industrialized AI for maintaining the system, and has not yet shipped anything for people building with it. 8 affordances·8 techniques
• Nord Design SystemNordhealth Nord is Nordhealth’s healthcare (veterinary/therapy) design system: 156+ Lit-based <nord-*> web components, CSS custom-property tokens, Tailwind CSS v4 integration with an n: variant prefix, two brands with dark and high-contrast variants. Its AI story is entirely consumption-side and unusually current for July 2026: a dedicated /ai/ section, /llms.txt plus a ~1M-token /llms-full.txt, a /raw/<path>.md markdown twin for every docs page, and, the standout here, two officially published Agent Skills served from a /.well-known/skills/ discovery endpoint following the Cloudflare Agent Skills Discovery RFC, installable with npx skills add https://nordhealth.design. The building side is invisible: the source repo is explicitly private, so no CLAUDE.md/AGENTS.md/cursor rules are observable, and the promised MCP server is still roadmap, not shipped. 7 affordances·6 techniques

Emerging · 1

• U.S. Web Design System (USWDS)U.S. General Services Administration (GSA) / Technology Transformation Services USWDS is the sharpest public-sector contrast case in this study: a heavily-used, actively-committed federal design system with essentially zero consumer-facing AI affordances and exactly one builder-facing one. There is no llms.txt, no llms-full.txt, no /ai docs page, no official MCP server, no Code Connect, no distributed editor rules. designsystem.digital.gov returns a 404 HTML page for every AI-convention path probed. What the team does ship, as of 2026-07-06, is a single official AGENTS.md at the repo root (PR #6748, merged), a terse, high-density operational brief for contributors’ coding agents that reads as pure onboarding compression rather than design-system coercion. Every agent affordance aimed at people building with USWDS is community-built and explicitly disclaims federal endorsement. 6 affordances·8 techniques