
Saglitzdesign Mcp
io.github.HalidSaglam由 GitHub 账号 HalidSaglam 发布v0.19.1更新于 Oct 24, 2026
Expert design & marketing MCP: UI/UX, design languages, SEO/GEO, tokens, a11y, recipes.
概览
本地 MCP 服务器,为编码助手提供精选的设计与营销知识库,以及设计令牌、配色、字体、动效生成和可访问性审计工具。
- 可以做什么
- 检索并阅读 96 篇设计、UX、营销、SEO 与 GEO 文档,获取各平台设计语言参考
- 生成设计系统:配色、字号阶梯、令牌、阴影、布局、动效与字体搭配,可导出为 CSS、Tailwind、SwiftUI、Compose 或 DTCG
- 获取组件规范、生产级参考代码与真实案例模式
- 运行确定性审计:WCAG 对比度与点击区域、设计 lint、设计系统与 UX 文案检查
- 生成评审清单与分阶段路线图,并对比 iOS、Android、macOS 与 Web 的规范差异
- 导入已有设计令牌并重新输出到其他平台
- 适用场景
- 适合助手在构建或评审 Web、iOS、Android、macOS 界面时使用,需要具体可执行的规则而非泛泛建议:落地页、仪表盘、移动端页面、付费墙、设计系统审计或跨平台移植。
- 运行要求
- 需要 Node.js 20 或更高版本(18 可运行但会警告)。从 npm 安装,由 MCP 客户端通过 stdio 启动;无需账号、API 密钥或网络访问。可选环境变量 SAGLITZDESIGN_KNOWLEDGE_DIR 可指向自有的 Markdown 规则目录。
安装
在 SourceWeft 中
- 打开 控制台中的 Saglitzdesign Mcp,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
SaglitzDesign MCP
An expert design & marketing brain for your AI coding agent.
A Model Context Protocol server that gives Claude, Cursor, and any MCP client expert‑level guidance on web, iOS, Android and macOS design — plus the UX, copywriting, SEO, GEO and marketing knowledge that makes a product actually convert.
96 curated knowledge documents · 36 tools · 9 build/review/port workflows · MCP resources with id autocomplete · a one‑call design‑system builder · real token/color/type/elevation/motion/a11y generators · design & UX‑copy linters · production component recipes · phased roadmaps · real‑world visual examples
[npm] [CI] [skills] [license] [MCP] [Glama]
Why · What's inside · Tools · Resources · Install · Usage · Your own rules · Changelog · License
Why
LLMs are confidently wrong about design. They reach for defaults, invent outdated specs, and give generic "make it clean and modern" advice. SaglitzDesign replaces that with a curated, sourced, prescriptive knowledge base your agent can query on demand — the kind of guidance you'd get from a senior product designer, a conversion copywriter, and a technical SEO all at once.
Ask your agent to design a paywall, audit a landing page, or plan an iOS app, and it pulls concrete rules ("44pt minimum touch target", "one primary button per screen", "LCP ≤ 2.5s"), real patterns from top apps, and a phased roadmap — instead of guessing.
- Runtime‑independent. The server reads only local files. No external API, no account, nothing to configure. It just works, offline.
- Local by design. It speaks MCP over stdio only and runs as a child process of your MCP client. There is no HTTP or SSE transport, so it cannot be hosted remotely or reached over a network — which is also why your prompts and code never leave your machine.
- Prescriptive, not vague. Every doc is written as rules an agent applies verbatim — numbers, thresholds, do/don't lists, anti‑patterns.
- Grounded. Design‑language specs from official sources; patterns studied from real top apps and sites; classics distilled from the actual books.
What's inside
96 knowledge documents across 11 categories:
Which documents' sources are enforced, and which are not. Of the 96
documents, 11 have their sources: checked by the test suite against a tiered
allowlist — the six Apple design‑language guides (apple-hig-liquid-glass,
ios-app-design, macos-app-design, apple-accessibility,
apple-shipping-readiness, wwdc-design-principles) and the five security
documents. The other 85 are curated but not yet checked, and extending the
assertion to them is its own piece of work rather than a formality: measured
today, 66 of the 85 would fail on 282 sources between them — 227 URL citations
across 133 distinct hosts, plus 55 that are not URLs at all and so have no
host to check. (The 227 and the 55 are the two halves of the 282, not two
figures to add.) Note
that apple-intelligence-design and visionos-spatial-design are Apple‑topic
documents that sit outside the enforced set. get_design_doc prints which
side of this line a document is on, next to its sources, so you never have to
come back here to find out.
Workflows (/ prompts) — "build me a…"
Beyond answering questions, SaglitzDesign ships prompts that orchestrate an
entire build end‑to‑end. In Claude Code they appear in the / menu under a name
that depends on how you installed it. Installed as the Claude Code plugin they
are /saglitzdesign:build_landing_page and so on — plugin commands, one
generated file per workflow in commands/. With the server installed on its own
they are /mcp__saglitzdesign__build_landing_page, which the menu labels
saglitzdesign:build_landing_page (MCP). That label is typeable too, (MCP)
suffix and all — it is just a clumsier way to reach the same prompt. What does
not work is that short form with the suffix dropped: Claude Code hands the bare
server:prompt alias to first‑party Anthropic connectors only — the URL has to
be https, on api.anthropic.com, under /v1/design/ — and no other server gets
one, stdio or remote. (The plugin's /saglitzdesign:build_landing_page above
only looks like that alias — it is a plugin command namespace, a different
mechanism.)
Invoke one and the agent runs the full method — roadmap → positioning & copy → generates
the design system (color, type, layout, elevation, tokens) → real examples →
writes the actual code from the component recipes → runs the deterministic
verify gate (design_lint, audit_accessibility, audit_design_system,
audit_ux_copy) → opens it in a browser, screenshots, scores it against the
critique rubric, and iterates until it passes.
Just type, e.g.,
/saglitzdesign:build_landing_page a SaaS invoicing tool for freelancers— everything after the command name becomes the brief, and the workflow asks for anything missing, then builds it. The/mcp__saglitzdesign__…form takes only the first whitespace‑separated word as its brief, so use the plugin command whenever the brief is a phrase. These are yours to type: every command carriesdisable-model-invocation, so the agent never starts one on its own. Asked in prose it reaches for the skills instead, and they cover much of the same ground — design review, landing‑page conversion, Apple platforms — as guidance rather than as this orchestrated build. Thesaglitzdesignskill's own trigger vocabulary names paywalls, so prose about one does reach a skill; what no skill carries isreview_paywall's RevenueCat benchmark scoring, which runs only when you type it.The visual critique loop uses whatever browser tool is connected (Claude in Chrome, Playwright, or chrome‑devtools MCP) to see and refine its own output. Without one, it reviews the code directly.
Tools
Resources
Beyond tools, the knowledge base is exposed as MCP resources, so clients that
support them (Claude Desktop, Cursor) can @‑mention a document directly —
with id autocompletion — instead of spending a tool call:
Visual examples
get_design_examples serves a curated library of real app/site screens. The
screenshots themselves are third‑party assets and are not redistributed, so
the published npm package ships the annotations and source links without the
images — the tool detects this and says so rather than pretending otherwise. If
you clone the repo you can rebuild the local image library; see
scripts/regenerate-examples.md.
Install
Requirements: Node.js 20+.
From npm (no clone needed):
From source:
Claude Code
Register it once (via npm — no clone), available in every project:
Or, if you cloned the repo, point it at the built file:
As a plugin
One install brings all three pieces — the MCP server, all nine skills (the eight depth skills plus the umbrella that routes into them) and a slash command for every workflow:
The plugin declares its server as npx -y saglitzdesign-mcp@latest, so the
server itself still comes from npm; the skills and the commands are files inside
the plugin. claude plugin details saglitzdesign@saglitz lists what arrived.
The three ways of installing do not carry the same payload:
The npm package's files: list covers dist/, knowledge/ and recipes/, and
does not name skills/ or commands/ — so an MCP-only install has every tool
and every document and none of the skill or command files. The skills CLI is the
mirror image: it copies each SKILL.md into your agent and brings no server, so
the tools those skills point at are not there unless you also install one.
Cloning the repository gets the source of all of it — but not a runnable
server: dist/ is gitignored, so a clone needs npm install && npm run build
before node dist/index.js starts. Everything else in the table is a tracked
file and arrives with the clone.
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
Cursor / other MCP clients
Run npx -y saglitzdesign-mcp over stdio (or node /absolute/path/to/dist/index.js from a clone).
Requires Node 20 or newer. It also runs on Node 18 with a warning; below that the launcher exits with an explanation instead of a syntax error.
Your client launches the server — you don't start it yourself. Run by hand it will print its status and then wait on standard input for a client that never arrives, which looks like a hang but is the server working correctly.
As skills (no MCP server)
Prefer lightweight, self-contained skills? Install the condensed SaglitzDesign skills into any skills-compatible agent:
Nine skills. saglitzdesign is the umbrella: it carries no depth guidance
of its own, and routes the work into whichever of the other eight fits. Those
eight — clean-interface-design, landing-page-conversion, design-review,
motion-and-animation, apple-platform-design, android-platform-design,
design-system-audit, ship-quality-gate — are each standalone guidance that also points to the
full MCP for depth. See skills/.
Each skill is copied into your agent and its content hash pinned in
skills-lock.json, so an installed skill does not change when this repository
does. Re-run the command above to pick up a new skill or an edited one.
Dev & debug
Usage
Once connected, just talk to your agent naturally — it decides when to call the tools:
"Using saglitzdesign, plan the design of an iOS fitness app." →
get_design_roadmapreturns a 7‑phase plan with the docs to read at each step.
"Review my landing page for conversion with saglitzdesign." →
design_review_checklist(landing‑page / conversion) +get_design_examples.
"How should a primary button behave on mobile?" →
get_component_guidancereturns specs, states, labels and anti‑patterns.
"Show me real paywall examples." →
get_design_examplesreturns annotated screenshots.
"What's llms.txt and how do I set it up?" →
seo_geo_guide(GEO) returns the tactic with a ready‑to‑use example.
Your own design rules
Point the server at a directory of your own and your documents join the base — searchable, readable, and, if you ask, part of the review checklist:
Or in claude_desktop_config.json / any MCP client, alongside command and args:
Several directories are allowed, separated the way PATH is on your platform.
- Your rules lead. When a search is genuinely about something you documented, your document comes first — measured by how many of the query's terms it covers, so a short house‑rules file is not buried by a long reference.
- Same id replaces ours. A file with
id: buttonstakes over from the built‑in one. That is deliberate, and it is announced at startup rather than happening quietly. Wherever it is served it is marked as your team's document, so nothing of yours is ever quoted as though it were sourced platform guidance. review: [website, saas-web-app]in the frontmatter puts the document into those project types'design_review_checklist, ahead of the curated list — the difference between your rules being findable and your rules being enforced.
Editing files inside the installed package works until
npm updatedeletes them. Use the environment variable.
Adding to this repository
Drop a Markdown file anywhere under knowledge/ with frontmatter:
The server indexes every .md on startup — no rebuild needed for content
changes (just restart the server). A /refresh-knowledge command
(.claude/commands/) can re‑research stale docs with agents.
How the knowledge was built
Design‑language and SEO/GEO docs were researched from official documentation
and current sources (cited in each file's sources). Real‑world UI patterns
were studied from top apps and websites; the visual example library was curated
the same way. Classic design and marketing books were distilled into original,
prescriptive syntheses — no source text is reproduced.
On images: screenshot files are a local research asset. They are not
included in this repository or any published package. Without them,
get_design_examples gracefully degrades to descriptions plus source links.
To rebuild the local image library (or add your own examples), see
scripts/regenerate-examples.md.
Find it on
[saglitzdesign-mcp MCP server]Also listed in the official MCP Registry and on npm.
License
MIT © 2026 Saglitz Design.
The knowledge/ documents are original syntheses with sources cited per file.
Referenced screenshots are not part of this repo and may not be redistributed —
see NOTICE.md.
来源:README.md,提交 b69521e
工具
0版本历史
1- v0.19.1最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分