A white-label design system in a box. Clone it, feed it a brand, and generate on-brand product UI for web, mobile and native, with AI agents that stay on brand because everything they touch is themed from the same tokens.
Nothing here is tied to any one company. Fork it, brand it, generate.
Prism System is a monorepo that carries everything a product team needs to look like itself: one token source of truth, a themed component library, build pipelines for web and iOS, a visual controller that turns a brand into a working theme in minutes, and an agent layer that generates product screens from that theme.
The mental model is a bubble you step inside. You run the controller, feed it your brand (colours, fonts, a logo, reference imagery, even a Figma file or an existing website), and it derives a complete token system and rebuilds every output so the whole repo now speaks your brand. From then on, any UI generated inside the repo is on brand by construction, because it is built from those tokens and nothing else.
| Layer | What it is | Count |
|---|---|---|
| Skills | Task recipes the agents follow, in packages/tokens/.agents/skills/ |
107 across 8 categories |
| Agents | Personas that own the work | 9 |
| Routes | The entry classifier that picks a skill sequence | 5 |
| MCP tools | The engine: token lookup, contrast, scoring, gating, audit | 24 |
| Packages | Tokens, three built themes, two component sets, icons | 8 |
| Platform outputs | CSS variables, SCSS, Swift | 3 |
A request flows top to bottom. The router is a small keyword classifier that picks one of five routes; each route names a fixed skill sequence; skills are run by agents; the whole thing sits on the MCP engine and ends in a generated artefact.
Entry /design-start · /engineer-start (one command per editor)
|
Router prototype · ui-craft · new-experience · handoff · corpus-distill
| (keyword classifier -> a fixed skill sequence per route)
|
Skills (107) 8 categories
| foundations 8 · design 18 · handoff 6 · react 20
| figma 21 · discovery 9 · quality 14 · workflow 11
|
Agents (9) Architect · Code · Content · Design Engineer · Designer
| Figma Executor · PM · React Expert · Testing
|
Engine MCP server, 24 tools (token-lookup, contrast, design-score,
| foundation-gate, audit, ramp-lookup, ...)
|
Artefact a themed screen or component under sandbox/<run>/
The registry lives in packages/tokens/.agents/skills/skills.json as a lean six-field index (name, description, category, path, agents, autonomy). Optional metadata (intents, the compose graph, inputs and outputs, gates) sits alongside it in skills.extended.json, so the file you scan day to day stays readable. Agents are defined in .github/agents/; the router is packages/tokens/.agents/skills/design/design-router.md with a fast heuristic in packages/tokens/mcp-server/tools/design-route.js.
The system is built to be forked and rebranded, not to serve one brand. Everything brand-specific is data, and everything generic is code. A new brand touches the token sources and the brand corpus, never the components, the skills or the pipelines. The same repo can carry a completely different brand by feeding the controller a different set of inputs and pressing Save.
Because the outputs are plain tokens (CSS variables, SCSS, Swift) and a standard component set, you can also lift the built theme into an app that lives outside this repo. Use the tokens or the @ds/ui components directly, or generate screens here and copy them out. Nothing locks you in.
One source, every platform. The controller writes your brand into packages/tokens/src/tokens.json and packages/tokens/data/resolved-hexes.json. A reconcile step in packages/output merges the two, then Style Dictionary builds the platform outputs: CSS variables into packages/theme-css, SCSS into packages/theme-scss, Swift into packages/theme-ios. The same build regenerates packages/ui/src/styles/theme.css, so the shadcn/ui components and everything built from them pick up the brand automatically.
The engine does real colour work, not just find-and-replace. From your seed colours it generates full perceptual ramps in OKLCH, checks WCAG contrast, and derives the semantic theme tokens (background, card, primary, muted, borders, the chart slots) for both light and dark. A material layer sits on top for surface style: flat, elevated, glass, soft or outline, with fine control over blur, opacity, tint and highlight, so you can dial in the iOS-glass look or keep it flat.
The components are the vendored shadcn/ui set in @ds/ui, with React components in @ds/components-react and Phosphor icons in @ds/icons, all themed from your tokens. To browse them, build the workspace packages once, then open Storybook:
npm run build # builds the packages
npm run storybook # http://localhost:6006Storybook documents the components alongside the token foundations, the colour palette and the type scale.
The controller is a zero-dependency Node server at tools/controller, served at http://localhost:4400. The left rail defines the brand; the right side previews it live against real surfaces.
The rail works top down. Theme presets give a one-click starting point (42 tweakcn themes) you then refine. Colours sets the primary, neutral tint, tertiary, semantic and chart hues, with a hue/saturation/brightness pass over the whole palette. Typography picks heading and body fonts, Shape controls radius, Spacing the density, Effects the shadows, and Material the surface style. A derived-tokens table shows every token the engine computes, so nothing that lands in the codebase is a mystery.
The preview has tabs so you can judge the theme honestly: a Moodboard, a full Dashboard, a Mail client, the Components sheet, Type specimens, and a Colours tab that lays out every generated ramp and the resolved theme tokens for the current mode.
Beyond colours and type, the controller captures your actual design language. Drop a reference screen or paste a Figma link and it reads the real specifics, spacing, radius, colour usage, component recipes and signature moves, straight from the Figma desktop MCP with no token or API key. You pick which values seed the system, and the generation agents read the captured language before composing any screen, so output looks designed rather than default. Identity records the name, tagline, voice and values; Guidelines collects the written rules. Every save takes a timestamped backup, so experiments are cheap.
git clone <your-repo-url> prism-system
cd prism-system
npm install --legacy-peer-deps
npm run controller # http://localhost:4400Set your brand. Pick a preset or set your colours, choose fonts, dial in shape, spacing and material. If you have a Figma file or a reference screen, capture its design language in the References tab and apply the values you want. Press Save & apply. The controller rewrites the token sources and rebuilds every output, so the whole repo now carries your brand.
Open the repo in your editor. The agent layer is shared across Claude, Cursor, Copilot, Codex and Continue, all reading the same skills. Kick off work with a command:
/design-start guided design intake: explore, create, check or hand off
/engineer-start turn a design or prototype into production code
The router classifies your request and loads only the skills that route needs, so context stays small and focused.
Describe what you want to build. Because everything is themed from the same tokens, the output is on brand automatically:
- Web UI comes out as
@ds/ui(the themed shadcn set) with icons from@ds/icons, bound entirely to tokens. - Mobile web uses the same components and the responsive token scale.
- Native iOS pulls the Swift tokens from
packages/theme-ios, so a SwiftUI screen matches the web to the pixel value.
Every run lands in a dated folder under sandbox/, with a manifest of what was asked and what came out. Nothing overwrites anything, so you can generate freely and keep what works.
Colour maths goes wrong in ways nothing complains about. A focus ring can sit at an opacity nobody can see, a stylesheet can claim one value and paint another, and every test still passes. So there is a battery you can run at any point:
npm run check # all eight
npm run check tokens gallery # just the ones you nameCheapest first. tokens recomputes every derived token from its own recipe and fails if the two disagree. gallery proves the preview is bound to tokens, with no stray literals. browser loads the generated theme into real Chromium and compares the pixel it paints with the hex the engine computed, for every derived token in both modes, which is the only honest way to know the resolved hex and the live color-mix() really are the same colour. behaviour and rail prove the controller's own layers are a no-op until you change something, so a fresh clone ships nothing you did not ask for. variables drives the Variables panel and checks that every token the system emits is explained there. audit renders the contrast audit for light and dark and fails on any pairing a derived state broke.
Six of those drive a real browser. Playwright is a devDependency so the package is already present, but the browser it drives is a separate install:
npx playwright install chromiumUntil you run that, the six browser checks skip rather than fail, and the summary tells you why. A missing binary is a setup fact, not a regression, and the run should not dress it up as six failures.
The tokens in packages/tokens/src are the single source of truth: when anything disagrees with them, they win. The brand, once saved, lives in the token sources and in design-corpus/brand/GUIDELINES.md and DESIGN-LANGUAGE.md. Generated work lands in sandbox/, one folder per run. START-HERE.md is the one-page orientation; design-start.md and engineer-start.md carry the two intake flows.
Node 22 or later and npm. The Figma desktop app with its Dev Mode MCP server enabled is optional, and only needed if you want the controller to read design language straight from a Figma file. The local claude CLI is optional too; without it, image analysis falls back to client-side colour extraction. Chromium is wanted only by the six browser checks: run npx playwright install chromium once and they run instead of skipping.
- The focus ring is visible now. It was a flat 30% wash of
--ring, which measured between 1.31:1 and 2.17:1 against the surfaces it lands on, well under the 3:1 that WCAG 2.2 asks of a focus indicator. It had been passing every test in the repo while being close to useless to anyone navigating by keyboard. It now solves its own opacity upward until the halo clears 3:1 on the page, the cards and the popovers, and it grades both emissions, because the resolved hex and the livecolor-mix()are genuinely different colours once they land on a card. Three places that were still mixing their own wash, at 25%, 30% and 50%, now read the token. - A derivation layer.
tools/controller/lib/derive.mjsturns the 34 chosen tokens into 96, of which 62 are derived. Every number in it is solved rather than declared: tier and hover mixes walk until the result is visibly distinct from what it came from, soft fills solve their alpha for a target ratio against the page, soft foregrounds walk until they clear AA on their own fill, and a hover is never allowed to break a pairing that passed before. Whatever the solver moved is reported, so the Variables panel shows its working instead of asking you to take a hex on trust. - Every derived token ships twice. A resolved hex for Style Dictionary, SCSS and Swift, and a live
color-mix()for the web and the controller preview. Thebrowsercheck exists to keep that claim honest. - Real gamut mapping. Colours that fall outside sRGB are now mapped with a binary chroma search against deltaEOK, the way CSS Color 4 specifies, rather than having their channels clipped. Ramps shift wherever a swatch used to be clipped,
warning light 450going from#fab053to#ffaa1aamong them. - The modal scrim dims in both modes. It washed the page in
--foreground, which in dark mode is near white, so it lifted the backdrop above the dialog sitting on it and inverted the depth cue a scrim exists to give. It now takes whichever pole of the page is actually the darker one. - Label inks flipped where they had to. Success, error and destructive were carrying white labels in light mode where a dark label reads better, and dark labels in dark mode where a light one does.
- The checks battery. Nine scripts under
tools/controller/checks/, wired tonpm run check, described above.
Prism System is public and open to contributions. Fork it, branch off main, and open a pull request. CONTRIBUTING.md covers the setup and the house rules (keep it brand-neutral, bind to tokens, never commit secrets), SECURITY.md explains how to report anything sensitive privately, and issue and pull-request templates live under .github/.
Prism System is open source and free. Use it for your own projects, or for work you do for your clients. It is not a product to be sold or resold; the value is in what you build with it, not in the system itself. Take it, brand it, ship real things with it. See LICENSE for the legal text.