MCP Elite Design Architect Hub
Enables multi-agent design sessions with a structured workflow for defining voice, generating directions, creating UI elements, and exporting artifacts.
README
MCP Elite Design Architect Hub
An advanced Model Context Protocol server plus a live frontend canvas for multi-agent design sessions.
What It Implements
- Mandatory
get_session_contexthandoff before any agent can mutate session state - Persistent session state with optimistic concurrency via
context_version - Strict staged workflow from voice definition through passes 1-5
- Radically distinct direction generation across spatial, typography, temperature, and motion axes
- Triple validation gate before UI elements are committed:
critique_current_designevaluate_design_qualitywow_factor_audit
- Snapshot and restore support
- Feedback memory that flags recurring issues
- Token mutation dependency alerts
read_session_statefor full canvas state readsread_user_suggestionsso agents can consume live user feedback pins- React/Tailwind/Motion export with zipped artifacts and Storybook stories
- WebSocket canvas bridge for real-time frontend updates
Projects
- MCP server: src/server.js
- Frontend canvas app: mcp-canvas-app
Install
npm install
Run The Canvas Server
node src/server.js
This starts:
- stdio MCP transport for tool calls
- WebSocket canvas bridge at
ws://127.0.0.1:3210/ws - health endpoint at
http://127.0.0.1:3210/health
Session state is persisted to .mcp-elite-design-state.json.
Run The MCP Stdio Server
node src/server-stdio.js
This entry point is for MCP clients such as Antigravity. It does not start the WebSocket or HTTP canvas bridge, so it avoids port 3210 entirely.
The canvas/live entry and the stdio entry are intentionally separate:
- Canvas/live server: src/server.js
- MCP stdio server: src/server-stdio.js
If someone still launches src/server.js in stdio mode, it now respects these environment overrides and will skip the live server:
MCP_TRANSPORT=stdioMCP_MODE=stdioMCP_DISABLE_LIVE_SERVER=1
Run The Canvas App
npm run dev --workspace mcp-canvas-app
One-Click Startup
Double-click start.bat from the project root.
It will:
- open the canvas server in one terminal window
- open the frontend dev server in a second terminal window
- wait for both services to respond
- open
http://localhost:5173in your default browser
Optional frontend env vars:
VITE_MCP_WS_URLdefault:ws://127.0.0.1:3210/wsVITE_SESSION_IDdefault:default-design-sessionVITE_CANVAS_AGENT_IDdefault:canvas-user
You can also open a specific session via query string:
http://localhost:5173/?session=my-session-id
Tool Flow
get_session_contextdefine_designer_voiceload_design_referencesgenerate_design_directionsapprove_design_directionapply_design_systemdefine_site_structureset_design_phasecreate_ui_elementread_user_suggestionsexport_artifacts
Antigravity MCP Config
Use the stdio entry point, not src/server.js.
Example config: mcp.antigravity.config.example.json
Equivalent args change:
{
"command": "node",
"args": [
"/absolute/path/to/cloned/repo/src/server-stdio.js"
]
}
Notes
load_design_referencessupportsprovided_analysisso the workflow can continue in offline or sandboxed environments.mutate_design_tokensupdates tokens and reports which generated elements are likely affected.export_artifactswrites generated files and a zip archive toartifacts/<session_id>/by default.- The canvas listens for live session updates over WebSocket, so agent mutations appear without a page refresh.
Deployment Guide
If you are cloning this repository and want to deploy the Design Hub for your own team to use remotely, follow these steps to host the frontend canvas and backend server online.
1. Deploy the Frontend Canvas (Vercel)
The easiest way to host the mcp-canvas-app is via Vercel:
- Push your cloned repository to your own GitHub account.
- Go to Vercel and create a New Project.
- Import your GitHub repository.
- In the Vercel project settings:
- Framework Preset: Select Vite.
- Root Directory: Set this to
mcp-canvas-app.
- Click Deploy. Vercel will generate a public URL for your canvas (e.g.,
https://your-canvas.vercel.app).
2. Deploy the Backend MCP Server (Free Options)
The backend must run continuously to process live WebSocket connections and provide the MCP endpoints. Here are completely free options you can use:
- Adaptable.io (Highly Recommended): Excellent free tier for Node.js apps. Extremely easy to deploy from GitHub, supports WebSockets, and doesn't require a credit card.
- Alwaysdata: Provides a free 100MB environment. Great if you want to avoid credit cards and just need a small space to run the server.
- Render (Free Tier): Render offers a free Web Service tier. Note: Free instances will spin down after 15 minutes of inactivity, which can cause a short delay the next time you use it.
- Glitch: A great fully free playground that can host Node.js WebSocket servers (also spins down when inactive).
- Fly.io: Requires a credit card to prevent abuse (but won't charge you). Their free allowance is very generous for running small, always-on WebSocket servers.
General Deployment Steps (e.g., on Render):
- Create an account and start a new Web Service.
- Connect your GitHub repository.
- Use the following configurations:
- Root Directory:
.(the main repository folder) - Build Command:
npm install - Start Command:
node src/server.js
- Root Directory:
- Deploy the service. Your provider will give you a backend URL (e.g.,
https://my-backend.onrender.com).
3. Connect the Frontend and Backend
Once both are deployed, you must tell your frontend where the backend WebSocket is hosted.
- Go to your frontend hosting provider (Vercel) > Settings > Environment Variables.
- Add a new variable:
- Name:
VITE_MCP_WS_URL - Value: Your backend's WebSocket URL. (Change
https://towss://— e.g.,wss://my-backend.onrender.com/ws).
- Name:
- Trigger a redeploy of your frontend so the environment variables take effect.
Now, anyone with your frontend URL can view the live canvas, and any remote MCP agents connected to your backend will update the interface in real-time!
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。