design-extract

Built by Manav Arya Singh (@Manavarya09)

Sponsors

Atlas Cloud — powers OpenAI-compatible --smart classification. For AI coding workflows, try the Atlas Cloud coding plan.

Bloome — deploys your AI clone to handle every audience conversation, running in the cloud, 24/7.

Try designlang on Bloome.

designlang points a headless browser at any URL and reads the design system off the live DOM. One command emits 17+ files — DTCG tokens, Tailwind config, shadcn theme, Figma variables, motion tokens, typed component anatomy, brand voice, page-intent labels, and a paste-ready prompt pack for v0 / Lovable / Cursor / Claude Artifacts.

It also goes where extractors don't: layout patterns, responsive behavior across 4 breakpoints, hover / focus / active states, WCAG contrast scoring, multi-page consistency, drift checks against a live source-of-truth, visual-diffs, and a shareable graded report card.

Quick start

npx designlang https://stripe.com                      # extract everything
npx designlang dna stripe.com                          # place it in the measured design space: neighbours + percentiles ← v13.1
npx designlang site stripe.com                         # whole-site: one canonical system + consistency grade ← v12.23
npx designlang fidelity stripe.com --clone localhost:3000  # score a clone vs the original (visual + motion) ← v12.24
npx designlang gallery                                 # static shareable gallery of measured clones      ← v12.24
npx designlang studio                                  # live token editor: edit, preview, export, share ← v12.19
npx designlang verify stripe.com                       # fidelity score: rebuild from tokens vs live ← v12.18
npx designlang pair stripe.com linear.app              # fuse two designs (visuals A × voice B)    ← v12.8
npx designlang brand stripe.com                        # full brand-guidelines book (13 chapters)  ← v12.7
npx designlang theme-swap stripe.com --primary "#ff4800"  # recolour around your brand        ← v12.6
npx designlang pack stripe.com                         # one polished design-system directory ← v12.4
npx designlang remix stripe.com --as cyberpunk         # restyle in another vocabulary       ← v12.3
npx designlang remix stripe.com --all                  # emit all 6 vocabs at once           ← v12.3
npx designlang grade https://stripe.com --badge        # report card + SVG badge             ← v12.2
npx designlang battle stripe.com vercel.com            # head-to-head graded fight           ← v12.2
npx designlang clone https://stripe.com                # working Next.js starter
npx designlang --full https://stripe.com               # screenshots + responsive + interactions

Drop a live design-score badge in any README:

![Design Score](https://designlang.app/badge/stripe.com.svg)

Watch it read — live Extraction Theatre (v13)

Don't take the output on faith — watch it happen. On designlang.app/watch you paste a URL and a real headless Chromium opens the page and reads its entire design system in real time: a split stage with the live browser on the left and the design system — palette, type, spacing, motion — assembling itself on the right as each token lifts off the page.

It's driven by the actual extraction (the browser's CDP screencast streamed frame-by-frame alongside the real token events), recorded so a shared link replays the exact run. No install, no account. Every gallery card can be watched the same way, and the gallery now spans 37 real-graded design systems.

Whole-site design system (site)

Most extractors read a single URL. designlang site crawls a site's canonical pages (home, pricing, docs, blog, about, product…) and synthesizes one de-duplicated system. Every token is elected by coverage — the share of pages that use it — so what's genuinely site-wide is separated from one-off, page-local choices. Near-identical colours are merged in OKLab. It's fully deterministic and free; no API key.

npx designlang site stripe.com --max-pages 8

You get, alongside the standard pack emitted from the canonical system:

Measured clone fidelity (fidelity + gallery)

Cloning tools all claim "pixel-perfect" — none of them measure it. designlang fidelity does. Point it at the original and your clone (often a local dev server) and it returns one honest number, both halves of the clone:

• Visual — full-page screenshots of each, pixel-diffed.
• Motion — extractMotion() on both, compared across feel, durations, easings, springs, keyframe kinds, scroll-linked motion, and choreography/ stagger. (Most clones reproduce static pixels and drop the motion entirely; this is where they lose.)

The two blend into a 0–100 score + letter grade, and — the part competitors don't ship — a ranked correction plan: the exact next edits that will raise the score, hardest-hitting first. Measure → fix → re-run until it converges.

npx designlang fidelity https://stripe.com --clone http://localhost:3000
npx designlang fidelity https://stripe.com --clone http://localhost:3000 --min 90   # CI gate

Generating the clone with designlang clone? Add --fidelity and it grades the clone's token basis against the live site the moment it's built — writing FIDELITY.md + a correction plan straight into the project, no separate step:

npx designlang clone https://stripe.com --fidelity

You get fidelity.md (score + motion table + correction plan), fidelity.json, a shareable fidelity-card.svg (88% · B · stripe.com), and a fidelity-diff.png loss heatmap.

Then publish them. designlang gallery scans your reports and builds a deployable static site — an index of score cards plus a permalink page per clone (each with an OG card, so a shared link unfurls the number).

npx designlang gallery --title "Our clones" --base-url https://clones.example.com

A measured design space (dna)

Every extractor on the market answers what a design uses. None of them answers where it sits. designlang grade returns a letter — but a letter with no reference frame is a thermometer with no scale on it.

designlang dna reduces a design to 30 deterministic features across five axes — colour, type, space, shape, motion — and ranks it against a corpus of real design systems:

npx designlang dna raycast.com

  Design DNA · https://raycast.com
  30 features, 100% measurable

  vs 8 systems (default)

  color    65th percentile
  type     56th percentile
  space    42th percentile
  shape    79th percentile
  motion   67th percentile

  Nearest
  0.14  https://railway.app
  0.15  https://linear.app
  0.19  https://notion.so

Distance is the mean absolute difference across the features both designs have, so 0.14 reads as "the average feature is 14% of its range apart" — a number you can reason about, unlike a Euclidean distance in 30 dimensions.

The report goes past similarity to what makes a design look the way it does: the features furthest from the middle of the corpus, in plain language ("corner radius — 100th percentile, far above the corpus").

Bring your own reference frame — your products, your competitors, a target look:

npx designlang dna-corpus acme.com acme.com/pricing competitor.com
npx designlang dna acme.com/new-page --corpus ./corpus.json

Three things it refuses to do, because a design score that overclaims is worse than no score at all:

• A missing measurement stays missing. Features the page never exposed are null, never a substituted midpoint, and distance skips them instead of comparing against an invented value.
• Every number carries its evidence. Distances report how many features backed them, percentiles name the corpus and its size, and the report warns when coverage was partial.
• The space is versioned. Feature order and normalization are frozen per vector version, and a corpus built by different rules is rejected rather than silently compared against.

Install

npm i -g designlang                         # global
designlang install-browser                  # one-time: the Chromium designlang drives
npx skills add Manavarya09/design-extract   # as an agent skill (40+ agents)

npm install no longer downloads a browser, so it works in CI, Docker and behind proxies. install-browser fetches the Chromium matching the bundled Playwright (add --with-deps on bare Linux). Without it, designlang falls back to an installed Google Chrome.

Use in Claude Code (plugin)

designlang ships as a Claude Code plugin — eleven slash commands that wrap the CLI. Install it from inside Claude Code:

/plugin marketplace add Manavarya09/design-extract
/plugin install designlang@designlang

The first command registers this repo as a plugin marketplace; the second installs the designlang plugin from it. Restart Claude Code if prompted, then the commands are available:

Prefer raw MCP tools? The CLI ships an MCP server that extracts live URLs:

claude mcp add designlang -- npx -y designlang mcp

extract_design(url) returns a job id; poll get_job_status, then pass the id to get_tokens, get_colors, get_typography, get_components, get_findings, compute_drift (against a baseline job) or export (dtcg, tailwind, shadcn, figma, css, design-md). Add --output-dir to also serve an extraction already on disk as resources.

Atlas Cloud for --smart

designlang --smart can now use Atlas Cloud through its OpenAI-compatible chat API, so low-confidence classifiers can stay zero-dependency while routing to Atlas-hosted models.

export ATLASCLOUD_API_KEY=""
export ATLASCLOUD_MODEL="deepseek-ai/deepseek-v4-pro"

npx designlang https://stripe.com --smart

• Atlas Cloud envs: ATLASCLOUD_API_KEY, ATLASCLOUD_MODEL, optional ATLASCLOUD_API_BASE
• Alias envs also work: ATLAS_CLOUD_API_KEY, ATLAS_CLOUD_MODEL, ATLAS_CLOUD_API_BASE
• Existing OPENAI_API_KEY and ANTHROPIC_API_KEY flows keep working unchanged

What you get

Each run writes 17+ files to ./design-extract-output/. The headline outputs:

Multi-platform (--platforms web,ios,android,flutter,wordpress,all) adds ios/, android/, flutter/, and a WordPress block theme. --emit-agent-rules adds Cursor / Claude Code / generic agent rule files.

Why designlang vs anything else

Other tools give you the paint. designlang reads the architecture:

• Layout system — grids, flex containers, container widths, gaps — not just tokens.
• Responsive — crawls 4 breakpoints and reports what changes (--responsive).
• Interaction states — programmatically hovers and focuses, captures the deltas (--interactions, --deep-interact).
• Motion language — durations, easing families, spring detection, scroll-linked flag, feel fingerprint (springy / smooth / mechanical / mixed).
• Runtime motion (--motion-runtime) — drives the page (load / scroll / hover / focus) and reads document.getAnimations() to capture what actually animates: real durations, choreography/stagger sequences, and scroll recipes (parallax / reveal / pin). Folded into *-motion-tokens.json and previewable live in the studio's Motion tab.
• Component anatomy — slot trees with variant × size × state matrices, emitted as typed .tsx.
• Brand voice — tone, pronoun posture, heading style, CTA verb inventory.
• Page intent + section roles — landing / pricing / docs etc., with semantic regions (hero, feature-grid, pricing-table, cta…).
• Multi-page consistency — auto-discovers canonical pages, reconciles shared vs per-route tokens.
• WCAG — every fg/bg pair scored, with a remediation palette suggesting nearest passing colors.
• Drift + lint + visual-diff — designlang drift, lint, visual-diff all CI-ready, exit non-zero on failure.
• Live-site sync — treat the deployed site as source of truth (designlang sync).
• MCP server — designlang mcp exposes tokens, regions, components, and contrast pairs to any MCP-aware agent.

designlang grade https://stripe.com         # ← v12.1: shareable report card
designlang clone https://stripe.com         # → working Next.js app
designlang apply https://stripe.com -d ./app   # auto-detect framework, write tokens
designlang brands stripe.com vercel.com linear.app   # N-brand matrix
designlang drift https://yourapp.com --tokens ./src/tokens.json
designlang lint ./src/tokens/design-tokens.json     # CI-ready linter
designlang visual-diff https://staging.app https://app   # single-file HTML diff
designlang mcp                              # stdio MCP server for Cursor / Claude Code
designlang doctor                           # sanity-check the local install

All features

Full CLI Reference

designlang  [options]

Options:
  -o, --out          Output directory (default: ./design-extract-output)
  -n, --name        Output file prefix (default: derived from URL)
  -w, --width         Viewport width (default: 1280)
  --height            Viewport height (default: 800)
  --wait              Wait after page load for SPAs (default: 0)
  --dark                  Also extract dark mode styles
  --depth              Internal pages to crawl (default: 0)
  --screenshots           Capture component screenshots
  --responsive            Capture at multiple breakpoints
  --interactions          Capture hover/focus/active states
  --deep-interact         Auto-interact pass (scroll, menus, modals, accordions, hover CTAs)
  --motion-runtime        Capture runtime motion via getAnimations() (durations, choreography, scroll recipes)
  --full                  Enable all captures (implies --deep-interact + --motion-runtime)
  --cookie    Cookies for authenticated pages (name=value)
  --cookie-file     Load cookies from JSON / storageState / Netscape cookies.txt
  --header    Custom headers (name:value)
  --user-agent        Override the browser User-Agent string
  --insecure              Ignore HTTPS/SSL certificate errors (self-signed, dev, proxies)
  --selector         Only extract from elements matching this CSS selector (e.g. ".pricing-card")
  --system-chrome         Use the system Chrome install instead of the bundled Chromium (skips 150MB download)
  --json                  Print full extraction as JSON to stdout (for piping into other tools)
  --framework       Only generate specific theme (react, shadcn)
  --platforms        Additional platforms: web,ios,android,flutter,wordpress,all (additive)
  --emit-agent-rules      Emit Cursor / Claude Code / CLAUDE.md / agents.md rule files
  --tokens-legacy         Emit pre-v7 flat design-tokens.json shape (backward compat)
  --no-history            Skip saving to history
  --verbose               Detailed progress output

Commands:
  apply                        Extract and apply design directly to your project
  clone                        Generate a working Next.js starter from extracted design
  score                        Rate design quality (7 categories, A-F, bar chart)
  grade                        Generate a shareable HTML Design Report Card (--format html|md|json|svg|all, --badge, --open)
  battle                Head-to-head graded battle card (--format html|md|json|all, --open)
  remix                        Restyle in another vocabulary (--as brutalist|swiss|art-deco|cyberpunk|soft-ui|editorial, --all, --list, --open)
  pack                         Bundle every output into one design-system directory (--with-clone, --open)
  theme-swap  --primary   Recolour around a new brand primary (--from, --format html|md|json|tokens|all, --open)
  brand                        Generate a full editorial brand-guidelines book (--format html|md|json|all, --open)
  pair                  Fuse two designs across 7 axes (--colors-from, --typography-from, --spacing-from, --shape-from, --motion-from, --voice-from, --components-from, --brand)
  watch                        Monitor for design changes on interval
  diff                  Compare two sites' design languages
  brands                   Multi-brand comparison matrix
  sync                         Sync local tokens with live site
  history                      View design change history
  install-browser                   Download the Chromium designlang drives (--with-deps on bare Linux)
  mcp                               Launch stdio MCP server: extract live URLs as jobs (--output-dir  also serves one on disk)
  lint                        (v9) Audit a local token file (.json/.css) — CI-ready
  drift  --tokens        (v9) Check local tokens for drift against a live site
  visual-diff        (v9) Side-by-side HTML diff of two URLs
  fidelity  --clone  Score a clone vs the original — visual pixel-diff + motion fidelity → one grade + ranked correction plan (--min, --motion-runtime)
  gallery [dir]                     Build a shareable static gallery of measured clones (--title, --base-url)

Example output

designlang https://vercel.com --full →

Colors: 27 · Fonts: Geist + Geist Mono · Spacing: 18 (base 2px)
Shadows: 11 · Radii: 10 · CSS vars: 407 · Layout: 55 grids / 492 flex
Responsive: 4 viewports, 3 breakpoint changes · Interactions: 8 transitions
A11y: 94% WCAG · Score: 68/100 (D) · 4 issues

→ 17 files written to ./design-extract-output/
→ Run `designlang grade https://vercel.com` for a shareable report card

How it works

1. Crawl — Headless Chromium via Playwright, waits for network idle and fonts
2. Extract — One page.evaluate() walks up to 5,000 DOM elements, collecting 25+ computed properties, inline SVGs, font sources, and image metadata
3. Process — 17 extractor modules parse, deduplicate, cluster, and classify the raw data
4. Format — 12+ formatter modules emit the output files
5. Score — Accessibility extractor calculates WCAG contrast ratios for all color pairs
6. Capture — Optional: screenshots, responsive viewport crawling, interaction state recording

Install Everywhere

designlang ships surfaces beyond the CLI:

Claude Code plugin

Drop designlang straight into Claude Code as a plugin. Every CLI command becomes a slash command:

/plugin install Manavarya09/design-extract

Then inside any Claude Code session:

Manifest: .claude-plugin/plugin.json · marketplace: .claude-plugin/marketplace.json · commands: commands/ · skills: skills/.

Agent skill (other ecosystems)

Works with Cursor, Codex, and 40+ AI coding agents via the skills ecosystem:

npx skills add Manavarya09/design-extract

In Cursor / Codex / etc., use /extract-design .

Contributing

See CONTRIBUTING.md. PRs welcome!

License

MIT - Manav Arya Singh !!