grocers-design-mcp
Exposes the Grocers Design System as queryable tools for coding agents, including design tokens, UI components, brand foundations, and a skill manifest.
README
grocers-design-mcp
MCP server exposing the Grocers Design System as queryable tools for coding agents.
Hosted on Fly.io. Auth is a single long-lived token embedded in the URL path — no OAuth, no client config beyond pasting the URL.
Live: https://grocers-design-mcp.fly.dev/<TOKEN>/mcp
Health (no auth): https://grocers-design-mcp.fly.dev/health
What it serves
- Design tokens — every CSS custom property from the Grocers DS (colors, spacing, radii, shadows, type scale, motion). 101 tokens indexed at boot.
- Semantic type classes —
t-h1,t-display,t-body,t-price, etc. (16 ready-to-apply utility classes). - UI components — JSX source from the
web,mobile, andadminUI kits. 44 components — Dashboard, KpiCard, OrdersScreen, OrderRow, CartSheet, MProductTile, etc. - Brand foundations — content rules (tone, casing, no-emoji), visual rules (red ≠ error, destructive uses ink + type-to-confirm), iconography (Lucide, 2px stroke, 24px grid).
- Skill manifest — one-paragraph brand brief for fast agent bootstrap.
Tools
| Tool | Purpose |
|---|---|
list_tokens({category?, group?}) |
Discover tokens by category/group |
get_token({name}) |
Full detail for one token (with -- prefix) |
list_semantic_classes() |
t-* pre-built typography classes |
list_components({kit?}) |
Components by kit (web/mobile/admin) |
get_component({name, kit?}) |
JSX source + metadata |
get_foundations({section?}) |
README sections: visual/content/iconography/… |
get_skill() |
The SKILL.md manifest |
search({query, limit?}) |
Full-text search with pt-BR↔en synonyms |
validate_snippet({snippet, language?}) |
Lint CSS/JSX against 11 brand rules |
suggest_component({intent, kit?, include_source?}) |
Intent → idiomatic component or pattern |
Lint rules in validate_snippet
| Rule | Severity | Catches |
|---|---|---|
no-emoji |
error | Any emoji in source (incl. composed sequences) |
no-hardcoded-brand-color |
warning | #ff4444, #ed1f1f, … → suggests exact token |
no-hardcoded-ink-color |
warning | #0b0119, #14081f, … → suggests exact token |
no-arbitrary-tailwind |
warning | bg-[#ff4444], p-[13px], rounded-[7px] |
red-is-not-error |
warning | Brand red near 'erro/error/invalid' → use danger |
destructive-uses-ink-not-red |
warning | Destructive cue + brand red → ink + type-to-confirm |
off-scale-radius |
warning | border-radius not in {4,6,10,14,18,24,pill} |
off-scale-spacing |
warning | padding/margin/gap not in 4px scale |
non-brand-font |
warning | font-family not Neue Montreal/Geist Mono |
heavy-cta-weight |
info | font-weight: 600+ in button context |
missing-tabular-nums |
info | R$ N without tabular-nums or .t-price |
off-scale-shadow |
info | box-shadow not via var(--shadow-*) |
Intent map in suggest_component
30 curated entries spanning all three kits plus 5 non-component patterns
(destructive-action, error-state, price-display, icon-usage,
tone-of-voice). Kit inference from natural language ("mobile", "painel",
"web") biases the ranking.
Using with Claude Code
Add to ~/.claude.json or project .mcp.json:
{
"mcpServers": {
"grocers-design": {
"type": "http",
"url": "https://grocers-design-mcp.fly.dev/<TOKEN>/mcp"
}
}
}
The token lives in fly secrets and rotates via
fly secrets set GROCERS_MCP_TOKEN=$(openssl rand -hex 32) -a grocers-design-mcp.
Quick example flow (agent building a Grocers admin screen)
1. get_foundations({section:"visual"}) → brand rules
2. suggest_component({intent:"tabela de pedidos no admin"})
→ admin/OrdersScreen
3. suggest_component({intent:"modal de cancelar pedido"})
→ pattern destructive-action
(ink + type-to-confirm)
4. get_component({name:"OrdersScreen", kit:"admin"})
→ JSX source
5. validate_snippet({snippet: <composed JSX>}) → catch drift before ship
Local dev
npm install
GROCERS_MCP_TOKEN="$(openssl rand -hex 32)" PORT=8080 npm run dev
# Health: http://localhost:8080/health
# MCP: http://localhost:8080/<TOKEN>/mcp
Architecture
- Hono +
@hono/node-serverfor HTTP. - @modelcontextprotocol/sdk for JSON-RPC / streamable HTTP transport,
stateless mode (fresh
McpServerper request — no session state). - Token-in-URL middleware:
timingSafeEqualagainstGROCERS_MCP_TOKEN; wrong token returns 404 (not 401 — hides existence). - The DS bundle ships inside the container at
bundle/and is indexed once at boot into in-memory maps (O(1) lookups). - Fly app runs 2 machines (HA) in
gru, 256MB each, auto-stops when idle.
Deploy
fly deploy -a grocers-design-mcp --remote-only
Layout
bundle/ # Grocers DS handoff bundle (source of truth)
project/
colors_and_type.css # all design tokens
README.md # foundations
SKILL.md # one-paragraph brand brief
ui_kits/{web,mobile,admin}/ # React/JSX components
preview/*.html # visual specimens (not served by MCP)
src/
index.ts # entrypoint
server.ts # Hono app + token middleware
indexer.ts # CSS + JSX + README parser
tools/
tokens.ts # list_tokens, get_token, list_semantic_classes
components.ts # list_components, get_component
foundations.ts # get_foundations, get_skill
search.ts # search (synonym-aware)
validate.ts # validate_snippet (11 rules)
suggest.ts # suggest_component (30 entries)
Dockerfile
fly.toml
Roadmap
Out of scope for v0.3, candidates for next iteration:
- Multi-theme support (Zona Sul and other clients as override layers).
validate_snippet: semantic copy lint (e.g. "Erro: ..." string detection).suggest_component: broader UI primitive coverage (TableHeader, Pagination states, EmptyState variants).get_screen(name): return a parent component + all its children in one call.- shadcn/ui compatibility layer (map Grocers semantic tokens →
--primary/--backgroundetc.).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。