open-design-system-mcp-starter

A system-agnostic MCP server template for a design system. Point it at your own catalog, tokens and docs, and coding agents get a real API to ask instead of a memory to guess from: what components exist, what props they take, which token to write, whether a name they are about to use is actually real. A lookup miss never returns a fabricated answer; it returns "not in this system, nearest is X, because Y".

Quick start

1. Clone the repo.

2. Install dependencies:

   npm install

3. The repo ships a fictional example system, acme-elements, under examples/acme-elements, so the server runs immediately with no configuration:

   npm run serve

4. Try it. Register the example server in Claude Code:

   claude mcp add acme-elements -- npx tsx /absolute/path/to/src/cli.ts serve

   Then ask the agent "does this system have a Card component?" and expect a correction pointing at the nearest real element instead of an agent that invents one.

5. Point it at your own system next. When it is a checkout, run init --id <id> --root <path>; when it is only ever installed via npm, run init --id <id> --package <npm name> instead (see "Point it at your system" and "Systems consumed from npm" below).

Point it at your system

The recommended way to add a system is the init wizard. It scans a checkout, detects a catalog adapter, a tokens adapter and doc globs, and writes (or merges into) ds.config.json for you:

npx tsx src/cli.ts init

Answer its prompts, or skip them entirely with --id and --root:

npx tsx src/cli.ts init --id acme-elements --root ../acme-elements

Both forms accept --config <path> to target a config file other than ./ds.config.json. When only one of --id / --root is given, init still asks interactively, using the one you passed as that prompt's default.

When your system is only ever npm installed, never checked out, use package mode instead of --root: init --id <id> --package <npm name> [--foundations <npm name>] reads the installed package's own node_modules/<npm name> (barrel, compiled CSS, README) instead of scanning a source tree. See "Systems consumed from npm" below for the full walkthrough and what it detects.

To edit the config by hand instead, or to see every field init can set, the file that ships in the repo looks like this:

{
  "$schema": "./schema/ds.config.schema.json",
  "systems": {
    "acme-elements": {
      "name": "Acme Elements (example)",
      "description": "Placeholder system shipped with the starter so the server runs out of the box. Replace with your own.",
      "root": "examples/acme-elements",
      "componentModel": "custom-elements",
      "componentsPkg": "@acme/elements",
      "catalog": { "adapter": "custom-elements-manifest", "path": "custom-elements.json" },
      "tokens": { "adapter": "css-vars", "files": ["tokens.css"] },
      "docs": { "include": ["docs/**/*.md"] }
    }
  }
}

Every field of a system entry, from SystemConfig in src/types.ts:

Once the config points at your system, run:

npx tsx src/cli.ts extract --system <id>
npx tsx src/cli.ts doctor
npx tsx src/cli.ts serve

extract reads your source through the configured adapters and writes the ground-truth files into the system's data dir (data/<id>/catalog.json, tokens.json, docs-index.json, plus aliases.json when you maintain one). That data dir is committed on purpose: it is the server's ground truth, and a diff to it should be reviewed like a diff to code, not treated as build output. Pass --out <dir> to extract to write into a different directory instead, for a dry run or a CI check.

Every command accepts --config <path> (default ./ds.config.json); extract and doctor accept --system <id> to limit the run to one system.

Each extracted file is stamped with a hash of the source it came from. doctor compares that stamp against the current source and reports whether the catalog, tokens, docs and (when codeConnect is configured) the Code Connect mappings are each fresh. serve runs the same check on startup: it warns when any of them is older than the source, and refuses to start with --strict until you re-run extract.

Adapters

Adapters are the only place format knowledge lives. Everything downstream (search, tools, resources) reads the same SystemCatalog / SystemTokens shape regardless of which adapter produced it.

react-docgen

{ "adapter": "react-docgen", "src": "<dir relative to root>", "barrel": "<optional, relative to root>" }

The public API is whatever the barrel exports, not whatever docgen manages to document: named re-exports (export { A, B as C } from './x', including export { default as X }), export * from './x' followed recursively through nested barrels up to 6 hops, and local export const/function/class. A barrel written against the compiled output (export * from './core/index.js') resolves fine: .js/.jsx/.mjs/.cjs specifiers are matched back to their .ts/.tsx source before being read. barrel defaults to <src>/index.ts or index.tsx, or index.d.ts when neither exists, since a published package installed into node_modules ships only declaration files.

Props come from react-docgen-typescript, run once over every .tsx file under src, plus every .d.ts file so a published package's own declaration files are documented too. Literal union types are expanded into their member values rather than left as an alias name. A prop declared outside the system's own source, such as an inherited DOM attribute or a spread from a third-party base component, is recorded by name only, under inheritedProps, not given a full entry. A tsconfig compiler option this project's pinned TypeScript does not recognize (written for a newer TypeScript than the adapter runs on) is warned about and skipped rather than treated as a hard failure. When a barrel re-exports components (always true for a published package's dist/index.d.ts), the same component would otherwise be documented twice, once from its own file and once from the barrel; the adapter keeps one entry per name, and the component's own file wins over the barrel's re-export.

Props-type fallback. react-docgen-typescript's component detection is name- and shape-based, so it misses some perfectly ordinary components: a callable object (declare const Stack: { (props): JSX.Element }), or a props type that is an intersection with a union. For every barrel export docgen left with no props at all, the adapter reads the export's own <Name>Props (or <Name>BaseProps) type straight off a TypeScript checker built over the exports' declaring files, or, when there is no such type, the first parameter type of the export's call signature. An export documented this way carries docSource: 'props-type' instead of the default (docgen). A prop resolved this way whose declaration lives outside the system's own source is still recorded by name only, under inheritedProps, exactly like a docgen-documented export - an inherited DOM prop is name-only either way. extract prints a one-line summary of how many exports this closed the gap for, e.g. [extract] props-type fallback documented 12 of 34 undocumented exports. What is left after that line is what "Known limits" below, and the catalog overlay, are for.

css-vars

{ "adapter": "css-vars", "files": ["<file relative to root>", "..."] }

Parses --name: value; custom-property declarations out of plain CSS. :root/html/body selectors are the default ("light") theme; [data-theme="x"], .x-theme and .dark/.light classes, and @media (prefers-color-scheme: dark) are recognized as other themes. Tailwind 4 @theme blocks are read as default-theme declarations, and @layer wrappers (@layer utilities { :root { ... } }) are unwrapped and their contents processed normally. A brace-less at-rule ahead of a block, such as @import "./x.css"; or Tailwind's @custom-variant dark (...);, is ignored rather than folded into the next selector's header (which would otherwise mislabel that block's theme).

dtcg

{ "adapter": "dtcg", "files": ["tokens.json"] }

Parses Design Tokens Community Group JSON: $type is inherited from the nearest ancestor group when a token doesn't declare its own, {value} references are captured, and when multiple files are given each is matched to a theme (light/dark, or high-contrast) by its basename.

markdown-docs

docs.include (and optional docs.exclude) globs, relative to root, are chunked by heading into data/<id>/docs-index.json: one chunk per heading section, truncated to 1200 characters, each recording which catalog exports it mentions. This index is what get_guidance searches, and it is checked for freshness the same way the catalog and tokens are.

A glob walk skips node_modules and dot-directories, so it never reaches a README shipped inside an installed package. An entry in docs.include with no glob characters (*, ?, {, [) is taken as a literal path instead and read directly, bypassing that skip -- this is how a system consumed from npm points at a package's own README.md under node_modules.

Known limits

Neither docgen nor the props-type fallback above can document every component: some polymorphic or factory-built component (one assembled at runtime rather than declared as a plain function, for example) resolves to no props type and no callable signature the checker can read. Such a component still appears in allExports and still resolves through resolve_component as a real, existing symbol; it just has no props table. overlay-scaffold lists exactly which exports are still in this state and gives a team a starting point to fill them in by hand; see "Catalog overlay" below.

Catalog overlay

data/<id>/overlay.json holds hand-written facts merged over the extracted catalog at load time: for exports neither docgen nor the props-type fallback could document, or simply to add an example, a11y note, deprecation, or Figma link the source has no place to author. An overlay entry can only enrich a symbol that already exists in the catalog's allExports -- it never invents one, so a typo reports as unknown instead of silently creating a fake export.

{
  "exports": [
    {
      "displayName": "AcmeStack",
      "description": "Vertical layout primitive with a configurable gap.",
      "props": [{ "name": "gap", "type": "string", "required": false, "defaultValue": "\"md\"" }],
      "examples": [{ "title": "Basic stack", "code": "<AcmeStack gap=\"lg\">...</AcmeStack>", "language": "tsx" }],
      "a11y": { "accessibleName": "none" },
      "guidance": { "do": ["Use for vertical rhythm between blocks."], "dont": ["Don't use for inline layout."] },
      "deprecated": { "since": "2.0", "replacement": "AcmeStackV2" },
      "figma": { "nodeId": "1:23", "url": "https://figma.com/file/...?node-id=1-23" }
    }
  ]
}

Every field is optional except displayName. Precedence: an overlay props entry is merged into the extracted props list by name -- an overlay prop with the same name replaces the extracted one, any other extracted prop is kept, and a name the extractor never found is added. Every other field (description, a11y, guidance, deprecated, figma, tagName, inheritedProps, events, slots) is a plain overwrite when the overlay sets it; examples and docs are appended to, de-duplicated by content. Setting props also stamps docSource: 'overlay' on that export, same as when the overlay creates a whole new entry for a name docgen never produced one for.

unknown names -- a displayName the overlay lists that is not in allExports -- are never invented into existence. They are reported by doctor (overlay: N exports enriched, unknown names: ..., marked a failing check) so a typo or a renamed component gets caught instead of silently doing nothing.

The scaffold workflow:

npx tsx src/cli.ts overlay-scaffold --system acme-elements

writes data/<id>/overlay.json with one empty entry per undocumented export (the same set "Known limits" describes), each carrying a _note telling a human what to do: read the declaration file, fill in props (name, type, required, defaultValue, description), then delete the note. Re-running without --force only appends entries for names not already listed, so a team's edits to existing entries are never touched; --force rewrites the file from the current undocumented set.

Systems consumed from npm

No checkout is required. When your design system is only ever installed, not cloned, run the init wizard's package mode first:

npx tsx src/cli.ts init --id acme-elements --package @acme/elements

pointed at the app folder that has @acme/elements installed (default: the current directory). It reads node_modules/@acme/elements/package.json for the .d.ts barrel (types, typings, or exports['.'].types), finds compiled CSS from main/style/exports or by walking the package for a token-shaped stylesheet, picks up the package's own README.md, and writes the react-docgen + css-vars config below for you. Pass --foundations @acme/foundations when tokens ship from a separate package the components package does not declare as a peer dependency (when it does, the wizard finds it on its own). If @types/react is not installed in the app, the wizard prints a reminder: react-docgen-typescript needs it to resolve the package's declaration files.

Point the config at the app that has the packages installed instead of at the package's own repo, by hand, when the wizard's detection does not fit your package's layout, or to hand-tune what it wrote:

{
  "systems": {
    "acme-elements": {
      "root": "../my-app",
      "componentsPkg": "@acme/elements",
      "catalog": { "adapter": "react-docgen", "src": "node_modules/@acme/elements/dist", "barrel": "node_modules/@acme/elements/dist/index.d.ts" },
      "tokens": { "adapter": "css-vars", "files": ["node_modules/@acme/foundations/dist/tokens.css"] },
      "docs": { "include": ["node_modules/@acme/elements/README.md"] }
    }
  }
}

• root is the app folder that has the packages installed (its node_modules), not the design system's own checkout.

• catalog.src and catalog.barrel point into node_modules/<componentsPkg>/dist and its dist/index.d.ts; the react-docgen adapter falls back to index.d.ts for the barrel and reads .d.ts files alongside .tsx ones for exactly this case (see react-docgen above).

• tokens.files points at the foundations package's compiled CSS, e.g. node_modules/<foundationsPkg>/dist/tokens.css, when tokens ship from a separate package than the components.

• docs.include needs a literal file path, not a glob, to reach a README inside node_modules: globs skip node_modules by design, but a literal path is read directly regardless (see markdown-docs above).

• Install @types/react in the app so the components' declaration files resolve; without it, react-docgen-typescript cannot type-check the .d.ts files and documents nothing.

• When the package's barrel re-exports its components, as dist/index.d.ts normally does, the component's own declaration file wins over the barrel's re-export for the recorded props.

Aliases

Agents write the vocabulary they already know, not necessarily your system's vocabulary. Aliases are merged at load time from three layers, later layers winning on the same alias: team, then Code Connect, then the ecosystem lexicon.

• Ecosystem lexicon (src/data/convention-lexicon.json): names AI models tend to invent when a design system does not have them, mined from 898 graded agent generations across two production design systems. Never hand-edited; extended only with new mined evidence.

• Code Connect aliases (data/<id>/aliases.code-connect.json): written by extract when codeConnect is configured, from Figma enum property maps whose value differs from the code value (e.g. Figma's "Primary" versus the code's variant="primary"). See "Figma Code Connect" below. Not hand-maintained; regenerated on every extract.

• Team file (data/<id>/aliases.json): your own mapping from an alias to your system's real name. This is the file you maintain, and it wins over both of the above.

An alias without a team target still helps: the server can say "there is no Card here, the nearest concept is Surface" even before your team has mapped anything. A team entry that names a target turns that into a direct resolution.

Example data/acme-elements/aliases.json:

{
  "components": [
    {
      "alias": "Card",
      "concept": "surface-card",
      "target": "acme-surface",
      "source": "team"
    }
  ],
  "props": [
    {
      "alias": "spacing",
      "concept": "stack-gap",
      "target": "gap",
      "source": "team"
    }
  ]
}

Figma Code Connect

When your team already maintains Figma Code Connect files (figma.connect(Component, url, { props, example }) calls, usually in *.figma.tsx), point codeConnect at them and extract folds the mapping into the catalog and aliases:

{
  "systems": {
    "acme-elements": {
      "root": "../acme-elements",
      "catalog": { "adapter": "react-docgen", "src": "src" },
      "codeConnect": { "include": ["**/*.figma.tsx"] }
    }
  }
}

include (default **/*.figma.tsx, **/*.figma.ts) is a glob list, or a literal path, relative to root; set codeConnect.root instead when the mapping files live in a different checkout than the system itself (a separate storybook repo, say).

For every figma.connect(...) call that names a real catalog export, extract adds:

• A Figma node link -- figma: { nodeId, url, source } -- on the matching export, resolved from the connect call's URL. When docgen produced no CatalogExport for that name at all, a bare stub is created so the link still surfaces, the same way the catalog overlay does for an undocumented export.

• A Code Connect example -- the connect call's example callback, with its JSX pulled out and props.<name> occurrences rewritten to {<name>} placeholders -- appended to the export's examples.

• Component-value aliases -- for every props: { x: figma.enum('<FigmaValue>', { a: 'valueA' }) } map where a Figma-side value differs from the code-side value it maps to, an alias entry (source: 'code-connect') is written to data/<id>/aliases.code-connect.json. See "Aliases" above for how this merges: team beats Code Connect beats the lexicon on the same alias.

doctor reports Code Connect coverage when it is configured: code connect: N of M documented exports mapped (F files). Mappings are stamped and checked for freshness exactly like catalog, tokens and docs: a hash over the sorted list of matched files (paths and contents) is compared on every doctor and serve run, and serve --strict refuses to start when it is stale.

This does not replace Figma's own MCP server. Figma's MCP answers "what component is this node, with its current, live property values" straight from an open Figma file. This adapter answers a different question, offline: it reads the mapping files your team already wrote and maintains, and uses them to enrich this server's own catalog and alias data -- node links, a worked example, and vocabulary aliases -- without ever opening a Figma session itself.

Patterns

Nothing here is required to run the server, but a component-by-component API is not always enough: an agent asking "how do I compose a labeled field with an error" needs a recipe, not a props table. Patterns fill that gap, and get_pattern serves them.

Add markdown files to data/<id>/patterns/*.md. Each file is one pattern, with optional YAML-ish front matter between --- lines:

---
title: Labeled field with error
description: A single-line text field with its own label, ready to show a validation error.
tags: form, validation, error
language: html
---

# Labeled field with error

... prose ...

```html
<acme-text-field label="Email" invalid error-text="Enter a valid email address."></acme-text-field>
```

Front matter wins when present; anything missing falls back to the body: the first # heading becomes the title, the first paragraph becomes the description, and the first fenced code block becomes the code (its info string is the language, defaulting to tsx for a react system or html for custom-elements when neither front matter nor the fence say). Which catalog components a pattern uses is detected automatically from tags in the code, not authored by hand.

get_pattern searches these by plain-language query ("confirm dialog", "empty state") and returns the matched pattern's full prose and code. A system with no patterns/ directory yet gets a plain note back instead of an error, pointing at search_components in the meantime.

The example system ships three: confirm-actions.md, labeled-field.md and status-notice.md under data/acme-elements/patterns/.

Migration lookups

get_migration answers "is this deprecated, and what do I do instead" for a component or a specific prop (Button or Button.type), from three sources, in order:

1. The catalog's own deprecated field on the export or the prop, extracted mechanically by the catalog adapters -- this is the authoritative answer when present.

2. CHANGELOG*.md at the system's root (CHANGELOG.md preferred over any other CHANGELOG.*): lines that word-match the component or prop name, with the nearest preceding ##/### heading (usually the version) attached.

3. A codemods/ or scripts/codemods/ directory at the root: file names that mention the component are listed as available codemods; when a codemods directory exists but nothing matches, that is reported too, so an agent doesn't assume none exists.

Tools

Every tool returns a short text summary for the model plus a structuredContent object matching a type in src/types.ts. Nine tools ship:

No get_figma_mapping tool is planned: "which code component is this Figma node" is exactly what Figma's own MCP server already answers, with live property values from an open file. This starter's job is the catalog and validation side, which is why the Code Connect adapter feeds node links and aliases into the tools above instead of adding a tenth tool that would duplicate Figma's own.

Resources and prompts

Resources serve clients that attach context up front; tools serve agents mid-task. Both read the same data:

• ds://<system>/catalog

• ds://<system>/tokens

• ds://<system>/components/<name>

• ds://<system>/llms.txt

The build-ui prompt starts a UI task with the system's conventions and the tool routine already loaded, so an agent reaches for the right tool before it writes anything.

Registering the server

All registrations run the same command, with an absolute path to this repo's src/cli.ts and config:

npx tsx <abs>/src/cli.ts serve --config <abs>/ds.config.json

This is the stdio transport: one process per client, spawned by the client itself. For a team that wants one server everyone points at instead, see "Shared HTTP instance" below.

The repository ships a project-scoped .mcp.json that runs npx tsx src/cli.ts serve with a relative path, so opening a clone in Claude Code registers the example system with no setup. Edit the server name once you point the config at your own system, or delete the file if you register the server globally instead.

Claude Code

claude mcp add acme-elements -- npx tsx <abs>/src/cli.ts serve --config <abs>/ds.config.json

Or in .mcp.json:

{
  "mcpServers": {
    "acme-elements": {
      "command": "npx",
      "args": ["tsx", "<abs>/src/cli.ts", "serve", "--config", "<abs>/ds.config.json"]
    }
  }
}

Cursor

.cursor/mcp.json:

{
  "mcpServers": {
    "acme-elements": {
      "command": "npx",
      "args": ["tsx", "<abs>/src/cli.ts", "serve", "--config", "<abs>/ds.config.json"]
    }
  }
}

Codex

~/.codex/config.toml:

[mcp_servers.acme-elements]
command = "npx"
args = ["tsx", "<abs>/src/cli.ts", "serve", "--config", "<abs>/ds.config.json"]

Shared HTTP instance

For a team, run one server everyone points at instead of every client spawning its own stdio process: serve --http starts the same server over streamable HTTP.

npx tsx <abs>/src/cli.ts serve --config <abs>/ds.config.json --http --port 3333 --host 127.0.0.1 --path /mcp

• --port defaults to 3333, --path to /mcp.

• --host defaults to 127.0.0.1 (loopback only). Set it to 0.0.0.0 to expose the server beyond localhost, and set DS_MCP_TOKEN in the environment whenever you do -- the server prints a startup warning if it is bound to a non-loopback host with no token configured.

• DS_MCP_TOKEN, when set, is required as a bearer token on every request to --path (Authorization: Bearer <token>); with no token set, the endpoint is unauthenticated, same as the stdio transport.

• GET /healthz (no auth) returns each configured system's id, name, export and token counts, and catalog/tokens/docs freshness, plus the number of live sessions and the session limits -- point a load balancer's health check or a quick "is this instance current" glance at it.

• Sessions expire: a client that stops talking without a DELETE is evicted after 30 minutes idle (DS_MCP_SESSION_TTL_MS), and at most 200 sessions are held at once (DS_MCP_MAX_SESSIONS); when the cap is reached the least recently seen session is evicted first. Both are logged.

• Security: the token check is constant-time, but there is no rate limiting and no TLS. Put a reverse proxy with TLS in front of an exposed instance, keep --host 127.0.0.1 for a single machine, and treat the catalog as what it is: a description of your public component API, not a secret.

Registering a shared instance in Claude Code is a URL, not a spawned command (fictional host and id, use your own):

claude mcp add --transport http acme-elements http://ds-mcp.internal:3333/mcp

generate well-known --server-url <url> advertises the same URL alongside the stdio start command in the generated .well-known/mcp/servers.json, so a client that reads that file can reach either transport:

npx tsx src/cli.ts generate well-known --system acme-elements --target . --server-url http://ds-mcp.internal:3333/mcp

Generating the agent surface

extract produces the ground truth the server reads. generate writes the agent-facing surface derived from that data into an adopting team's own repo (their design system repo, or an app that consumes it) -- so an agent that hasn't attached this MCP server yet, or a plain-text client that can't call tools, still gets a grounded starting point.

npx tsx src/cli.ts generate <agents-md|llms-txt|skill|editor-rules|well-known|all> --system <id> --target <dir> [--force] [--server-url <url>]

--target is the repo to write into, defaulting to the current directory. --system picks the system when more than one is configured. --server-url only matters for well-known: it advertises a hosted streamable-http endpoint (a "Shared HTTP instance", see above) alongside the stdio command; omit it when the only registration you support is stdio.

Generated surface

Every generated section carries the same content: the system's name and consumption rule (import by name, or register-once-then-tag), the MCP server start command, the fixed 8-step tool routine (search_components -> resolve_component -> get_component -> find_token -> check_usage -> get_pattern -> get_guidance -> get_migration), the never-invent rule, any team aliases that resolve to a real target, and a few do/don't lines derived from the live catalog (a deprecated component or prop, a component whose accessible name is required, always calling find_token before a raw value) -- never a generic placeholder.

Two write rules, so generate is always safe to re-run as the catalog changes:

• Begin/end markers. AGENTS.md, CLAUDE.md and copilot-instructions.md mix hand-written and generated text. A re-run replaces only the text between <!-- ds-mcp:begin <id> --> and <!-- ds-mcp:end <id> --> for that system id; everything else in the file, including the team's own prose above and below, is untouched. AGENTS.md and copilot-instructions.md are created if missing; CLAUDE.md is only ever updated, never created, since not every team has one.

• Skip unless --force. The fully-generated files (llms.txt, llms-full.txt, both SKILL.md files, <id>.mdc, servers.json) carry an embedded signature marking them as written by this tool. A file already at that path is left alone -- and reported as skipped -- unless it carries that signature (a previous run of this tool) or --force is passed, so a team's own hand-written file at the same path is never silently clobbered.

generate llms-txt (and all) always reports the size of both files, e.g. llms.txt 1.4 KB, llms-full.txt 12.7 KB, so the context cost of attaching them is visible before anyone does.

Skills

The repo ships one skill for onboarding the starter itself: .claude/skills/ds-mcp-setup/. It walks a setup session through adding a system (init), extracting, running doctor, smoke-testing the server, generating the agent surface (generate all), registering the server with a client, and authoring aliases.json from what an agent got wrong on its first real session -- each step names a concrete check rather than "looks fine."

generate skill writes two more, generated per system rather than shipped: use-<id> (the build routine with concrete tool calls, this system's own aliases and do/don'ts) and review-<id> (how to review a diff against the system -- run check_usage per changed file, treat an unknown-component or invented-prop finding as blocking, and name the deprecated replacement instead of letting it pass silently). See "Generating the agent surface" above for where these land and when a re-run touches them.

Multi-kit systems

A design system with more than one consumable kit, for example a React kit and a web-component kit, declares one system per kit in ds.config.json rather than one system covering both. Every tool takes an optional system argument: omit it when only one system is configured, pass it to disambiguate when several are.

Development

npm run check        # typecheck + the offline test suite
npx tsx src/cli.ts extract && npx tsx src/cli.ts doctor   # the shipped example must stay idempotent and green

Node 22 or newer (.nvmrc pins 22). The GitHub Actions workflow in .github/workflows/ci.yml runs the same checks plus a git diff --exit-code -- data guard, so a change that alters the committed example data without meaning to fails CI. See CONTRIBUTING.md before opening a pull request.

Conventions: ESM TypeScript, strict, explicit .ts import extensions, no build step. Tests are offline (node:test via tsx, the MCP SDK's in-memory transport, synthetic data in temp dirs) and never call a live model or network endpoint. Nothing under src/ names a real design system, package or component; every name resolves from ds.config.json.

typescript is pinned to ^5 on purpose: TypeScript 7, the native rewrite, ships no classic Compiler API, so neither react-docgen-typescript nor this repo's own barrel parser can load under it. Keep typescript@^5 as the installed package, or if you need TypeScript 7 elsewhere in your workflow, install typescript@5 as a nested dependency of react-docgen-typescript instead of upgrading the top-level package.

Later

Pairing with a benchmark. The catalog and token contracts here are a superset of what open-design-system-bench extracts, so its output loads through the catalog-json adapter with no conversion. The starter does not depend on this; see docs/bench-pairing.md for the full note.