MCP Server Template
A skeleton template for building MCP servers with SEP-1865 interactive widget support, dual authentication (OAuth + API Key), and Cloudflare Workers deployment.
README
{{SERVER_NAME}} MCP Server
MCP Apps server with SEP-1865 interactive widget support, dual authentication (OAuth + API Key), and Cloudflare Workers deployment.
Quick Start
1. Replace Placeholders
Search and replace these placeholders in all files:
| Placeholder | Description | Example |
|---|---|---|
{{SERVER_NAME}} |
Human-readable name | "Currency Converter" |
{{SERVER_ID}} |
kebab-case identifier | "currency-converter" |
{{McpAgentClassName}} |
PascalCase class name | "CurrencyConverterMcp" |
{{SERVER_DESCRIPTION}} |
Brief description | "Convert currencies using real-time exchange rates" |
{{WIDGET_TITLE}} |
Widget HTML title | "Currency Converter Widget" |
{{GITHUB_ORG}} |
GitHub organization | "your-org" |
2. Update wrangler.jsonc
- Replace
{{SERVER_ID}}with your server ID - Replace
{{McpAgentClassName}}with your class name - Update the custom domain pattern
3. Install and Build
npm install
npm run build:widgets
4. Set Secrets
wrangler secret put WORKOS_CLIENT_ID
wrangler secret put WORKOS_API_KEY
5. Deploy
Deployment is automatic via Cloudflare Workers Builds when you push to GitHub.
For manual deployment (not recommended):
npm run deploy
Project Structure
src/
index.ts # Entry point with dual auth routing
server.ts # McpAgent class (OAuth path)
api-key-handler.ts # API key authentication with LRU cache
types.ts # Environment bindings
server-instructions.ts # LLM system prompt instructions
auth/
authkit-handler.ts # WorkOS OAuth with PKCE
apiKeys.ts # API key validation
auth-utils.ts # User lookup, HTML pages
props.ts # Auth context type
session-types.ts # Session interfaces
helpers/
assets.ts # loadHtml() for widget loading
resources/
ui-resources.ts # SEP-1865 UI resource definitions
tools/
descriptions.ts # Tool metadata (4-part pattern)
shared/
logger.ts # Structured logging
optional/ # Advanced features (delete if not needed)
web/
widgets/
widget.html # Widget entry point
widget.tsx # React widget component
components/ # shadcn/ui components
lib/ # Utilities (cn, types)
styles/ # Tailwind CSS
dist/widgets/ # Built output (gitignored)
Adding New Tools
When adding a tool, update these locations:
1. Tool Metadata (src/tools/descriptions.ts)
"your-tool": {
title: "Your Tool",
description: {
part1_purpose: "What it does...",
part2_returns: "Returns X, Y, Z...",
part3_useCase: "Use when...",
part4_constraints: "Note: limitations..."
},
examples: [...]
}
2. Server Registration (src/server.ts)
registerAppTool(
this.server,
"your-tool",
{
title: TOOL_METADATA["your-tool"].title,
description: getToolDescription("your-tool"),
inputSchema: { /* Zod schema */ },
_meta: { [RESOURCE_URI_META_KEY]: widgetResource.uri }
},
async (args) => { /* implementation */ }
);
3. API Key Handler (src/api-key-handler.ts)
Duplicate the tool registration for API key authentication path.
SEP-1865 MCP Apps Pattern
This skeleton uses the Two-Part Registration pattern:
- PART 1: Register Resource - UI HTML template from Assets
- PART 2: Register Tool - Links to resource via
_meta[RESOURCE_URI_META_KEY]
Data flows:
Tool Result -> structuredContent -> postMessage -> Widget State
Authentication
OAuth 2.1 (OAuth-capable clients)
- Flow:
/authorize-> WorkOS AuthKit ->/callback-> Tools - PKCE support (RFC 7636)
- Centralized login at panel.wtyczki.ai
- Example clients: Claude Desktop
API Key (Non-OAuth clients)
- Header:
Authorization: Bearer wtyk_xxxxx - Keys generated via panel.wtyczki.ai
- LRU cache prevents memory leaks
- Example clients: AnythingLLM, Cursor
Widget Development
Key Concepts
- React 18 with
useApp()hook from@modelcontextprotocol/ext-apps/react - Tailwind CSS with automatic dark mode (via host context)
- Fixed 600px height container (mandatory for MCP Apps)
- viteSingleFile inlines all JS/CSS into single HTML
Development
# Build widget
npm run build:widgets
# Watch mode
npm run dev:widget
# Full dev (server + widget watch)
npm run dev:full
Widget Lifecycle
const { app } = useApp({
onAppCreated: (app) => {
app.ontoolinput = (params) => { /* tool called */ };
app.ontoolresult = (result) => { /* display result */ };
app.onhostcontextchanged = (ctx) => { /* theme change */ };
app.onteardown = async () => { /* cleanup */ };
}
});
Configuration Files
| File | Purpose |
|---|---|
wrangler.jsonc |
Cloudflare Workers config (bindings, routes) |
vite.config.ts |
Widget build config |
package.json |
Dependencies and scripts |
tsconfig.json |
TypeScript config (server) |
web/tsconfig.json |
TypeScript config (widget) |
Environment Variables
Set via wrangler secret put:
| Variable | Required | Description |
|---|---|---|
WORKOS_CLIENT_ID |
Yes | WorkOS client ID |
WORKOS_API_KEY |
Yes | WorkOS API key (starts with sk_) |
AI_GATEWAY_TOKEN |
No | AI Gateway token (if using Workers AI) |
Common Issues
Widget not loading
- Check
npm run build:widgetscompleted successfully - Verify
web/dist/widgets/widget.htmlexists - Check ASSETS binding in wrangler.jsonc
Authentication failures
- Verify WORKOS_CLIENT_ID and WORKOS_API_KEY secrets are set
- Check USER_SESSIONS KV namespace is configured
- Ensure custom domain is set up in Cloudflare
Tool not appearing
- Check tool is registered in BOTH server.ts AND api-key-handler.ts
- Verify tool name matches exactly in all locations
- Check handleToolsList() includes the tool schema
Production Checklist
- [ ] All
{{PLACEHOLDER}}values replaced - [ ] wrangler.jsonc configured with correct IDs
- [ ] Secrets set via
wrangler secret put - [ ] Custom domain configured in Cloudflare
- [ ] GitHub repository connected to Cloudflare Workers Builds
- [ ] Widget builds successfully (
npm run build:widgets) - [ ] Type checking passes (
npm run type-check)
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。