MCP Demo Server
Model Context Protocol (MCP) server implementation demonstrating production-ready patterns in TypeScript, with tools for calculation, weather, and filesystem access.
README
MCP Demo Server
Model Context Protocol (MCP) server implementation demonstrating production-ready patterns in TypeScript. Four parallel implementations comparing bare-metal and framework approaches.
Overview
This project showcases MCP server implementations with:
- Three core patterns: Computation (calculate), API integration (weather), filesystem access
- Four implementations: Bare-metal, FastMCP, EasyMCP*, mcp-framework
- Framework comparison: Real-world evaluation of MCP development approaches
*Note: The EasyMCP implementation may not run due to unstable npm package (v0.0.0-development).
- Enterprise patterns: Security, observability, reliability demonstrated across all implementations
- Dual transports: Stdio for desktop apps, HTTP/SSE for remote access
Quick Start
Prerequisites
- Node.js 20+
- pnpm 8+
- OpenWeatherMap API key (free tier: https://openweathermap.org/api)
Installation
# Clone and install
git clone <repository-url>
cd mcp-demo
pnpm install
# Configure
cp .env.example .env
# Edit .env and add your WEATHER_API_KEY
Use with Claude Code
# Build the server
pnpm build
# Add to Claude Code
claude mcp add demo-server node -- $(pwd)/dist/servers/bare-metal/index.js
# Verify connection
claude mcp list
Manual Testing (HTTP mode)
# Development
TRANSPORT_MODE=http pnpm dev
# Production
pnpm build
TRANSPORT_MODE=http pnpm start
Note: For Claude Code (stdio mode), you don't need to manually run the server - it auto-starts!
Architecture
High-Level Design
┌─────────────┐
│ Client │
└──────┬──────┘
│
┌──────▼───────────────────┐
│ Transport Layer │
│ • Stdio (stdin/stdout) │
│ • HTTP/SSE (Express) │
└──────┬───────────────────┘
│
┌──────▼───────────────┐
│ Middleware Pipeline │
│ • Auth │
│ • Rate Limiting │
│ • Logging/Tracing │
│ • Metrics │
└──────┬───────────────┘
│
┌──────▼────────────┐
│ Message Router │
│ • tools/list │
│ • tools/call │
│ • resources/list │
│ • resources/read │
└──────┬────────────┘
│
┌──────▼──────────────────┐
│ Handler Registries │
│ • ToolRegistry │
│ • ResourceRegistry │
└──────┬──────────────────┘
│
┌──────▼────────────────────┐
│ Capability Handlers │
│ • CalculateTool │
│ • WeatherTool │
│ • FileResourceHandler │
└───────────────────────────┘
Handler Pattern
All capabilities implement a common interface:
interface CapabilityHandler<TParams, TResult> {
validate(input: unknown): ValidationResult<TParams>;
execute(params: TParams, context: ExecutionContext): Promise<Result<TResult>>;
handleError(error: Error): MCPError;
}
This eliminates repetition and makes adding new capabilities simple:
- Implement the interface
- Register with appropriate registry
- Done - routing and transport work automatically
Framework Comparison
This project includes four parallel implementations of the same MCP server capabilities, allowing you to compare approaches:
| Implementation | Lines of Code | Pattern | Best For |
|---|---|---|---|
Bare-metal (/servers/bare-metal) |
~2,000 | Direct MCP SDK | Learning MCP internals, full control, custom needs |
FastMCP (/servers/fastmcp-impl) |
~420 | Builder API | Production apps, Express-like familiarity |
EasyMCP (/servers/easymcp-impl) ⚠️ |
~400 | Decorators | Rapid prototyping, minimal boilerplate (package unstable) |
mcp-framework (/servers/mcp-framework-impl) |
~455 | Auto-discovery | Large projects with many capabilities |
⚠️ EasyMCP implementation may not run - the easy-mcp npm package has broken exports in v0.0.0-development
Feature Matrix
| Feature | Bare-metal | FastMCP | EasyMCP | mcp-framework |
|---|---|---|---|---|
| Transport: stdio | ✓ | ✓ | ✓ | ✓ |
| Transport: HTTP/SSE | ✓ | ✓ | ✗ | ✓ |
| Input validation | Manual Zod | Zod schemas | Type inference | Zod + helpers |
| Auto-discovery | ✗ | ✗ | ✗ | ✓ (from /tools) |
| Type safety | Manual | Schema-based | Decorator inference | MCPInput<this> |
| Setup complexity | High | Medium | Low | Medium |
| Boilerplate | High | Low | Minimal | Low-Medium |
| Framework dependency | None | FastMCP | EasyMCP | mcp-framework |
| Learning curve | Steep | Gentle | Gentle | Medium |
Code Comparison: Adding a Tool
Bare-metal (must implement full interface):
class MyTool implements CapabilityHandler {
validate(input: unknown): ValidationResult { /* ... */ }
execute(params: MyParams, ctx: ExecutionContext): Promise<Result> { /* ... */ }
handleError(error: Error): MCPError { /* ... */ }
}
// Register manually
toolRegistry.register('my_tool', new MyTool());
FastMCP (builder pattern):
server.addTool({
name: 'my_tool',
parameters: z.object({ input: z.string() }),
execute: async (args) => processInput(args.input),
});
EasyMCP (decorators):
class MyMCP extends EasyMCP {
@Tool({ description: 'Process input' })
async myTool(input: string) {
return processInput(input);
}
}
mcp-framework (auto-discovery):
// In /tools/my_tool.ts
class MyTool extends MCPTool {
name = 'my_tool';
schema = defineSchema({ input: z.string().describe('Input') });
async execute(input: MCPInput<this>) {
return processInput(input.input);
}
}
export default MyTool;
// Automatically discovered - no registration needed
When to Choose Each Approach
Choose Bare-metal when:
- Learning MCP protocol internals
- Need maximum control and customization
- Building something unusual or experimental
- Want minimal dependencies
- Performance optimization is critical
Choose FastMCP when:
- Building production applications quickly
- Team is familiar with Express.js patterns
- Need session management and authentication
- Want a proven, battle-tested framework
Choose EasyMCP when:
- ⚠️ Note: Currently not recommended - package has stability issues
- Rapid prototyping or MVPs
- Building simple, small servers
- Team prefers TypeScript decorators
- Want absolute minimum boilerplate
- Developer experience is top priority
Choose mcp-framework when:
- Building large projects with many capabilities
- Want CLI tooling (
mcp validate,mcp add) - Team prefers convention over configuration
- Scalability and organization are important
- Need structured file organization
Running the Implementations
Each implementation is in its own directory under /servers:
# Bare-metal
cd servers/bare-metal
pnpm build
node dist/index.js
# FastMCP
cd servers/fastmcp-impl
npm install
npm run dev
# EasyMCP (⚠️ may not work - package unstable)
cd servers/easymcp-impl
npm install
npm run dev
# mcp-framework
cd servers/mcp-framework-impl
npm install
npm run dev
All implementations support the same three capabilities (calculate, get_weather, file access) with identical APIs.
Capabilities
1. calculate (Computation Pattern)
Evaluates mathematical expressions with security controls.
Method: tools/call
Parameters:
{
"name": "calculate",
"arguments": {
"expression": "2 + 2 * 10"
}
}
Returns:
{
"expression": "2 + 2 * 10",
"result": 22
}
Security: Expression sanitization prevents code injection. Only mathematical operators allowed.
2. get_weather (API Integration Pattern)
Fetches current weather data from OpenWeatherMap.
Method: tools/call
Parameters:
{
"name": "get_weather",
"arguments": {
"location": "San Francisco"
}
}
Returns:
{
"location": "San Francisco, US",
"temperature": 18.5,
"conditions": "partly cloudy",
"timestamp": "2025-10-20T10:37:00Z"
}
Features: 10-minute result caching, configurable timeout, error handling for API failures.
3. file:// resources (Filesystem Pattern)
Provides safe read access to files in allowed directories.
List resources:
{
"method": "resources/list"
}
Read resource:
{
"method": "resources/read",
"params": {
"uri": "file:///tmp/example.txt"
}
}
Security: Path traversal protection, restricted to ALLOWED_FILE_PATHS, MIME type detection.
Configuration
All configuration via environment variables. See .env.example for details.
Key Variables
# Transport
TRANSPORT_MODE=stdio # or "http"
HTTP_PORT=3000 # HTTP mode only
# Weather API
WEATHER_API_KEY=your_key # Required for weather tool
WEATHER_CACHE_TTL=600 # Seconds (10 min default)
# Security
ALLOWED_FILE_PATHS=/tmp # Colon-separated paths
API_KEYS=key1,key2 # HTTP mode only
RATE_LIMIT_REQUESTS=60 # Requests/min per client
# Observability
LOG_LEVEL=info # debug|info|warn|error
LOG_FORMAT=pretty # json|pretty
Deployment
Local Development
# Stdio mode with auto-reload
pnpm dev
# HTTP mode with auto-reload
TRANSPORT_MODE=http pnpm dev
Docker
# Build image
docker build -t mcp-demo .
# Run container
docker run -p 3000:3000 \
-e WEATHER_API_KEY=your_key \
-e API_KEYS=your_api_key \
mcp-demo
MCP Client Integration
Claude Code CLI (recommended):
# Build the server first
pnpm build
# Add to Claude Code (stdio mode)
claude mcp add demo-server node -- /path/to/mcp-demo/dist/servers/bare-metal/index.js
# Verify it's connected
claude mcp list
# The server will auto-start when Claude Code needs it
# No need to manually run pnpm start!
Claude Desktop (stdio):
{
"mcpServers": {
"demo-server": {
"command": "node",
"args": ["/path/to/mcp-demo/dist/servers/bare-metal/index.js"]
}
}
}
HTTP Client:
curl -X POST http://localhost:3000/mcp \
-H "Content-Type: application/json" \
-H "Authorization: Bearer your_api_key" \
-d '{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "calculate",
"arguments": {"expression": "2 + 2"}
}
}'
Sharing Your Server
Stdio vs HTTP: When to Use Each
Stdio Mode (Local, Single User):
- ✅ Personal use with Claude Code/Desktop
- ✅ Secure (OS-level permissions only)
- ✅ No network exposure
- ✅ Auto-starts/stops with client
HTTP Mode (Shared, Multi-User):
- ✅ Share with other Claude Code users
- ✅ Remote access (deploy to cloud)
- ✅ Multiple simultaneous clients
- ✅ Testing with curl/Postman
- ✅ Browser-based integrations
Deployment for Sharing
Quick Test (ngrok):
# Terminal 1: Start server
TRANSPORT_MODE=http pnpm start
# Terminal 2: Create tunnel
ngrok http 3000
# Share URL: Others add with
# claude mcp add --transport http your-server https://abc123.ngrok.io/mcp
Production (Cloud):
# Deploy to Railway, Render, Fly.io, AWS, etc.
# Set environment:
# TRANSPORT_MODE=http
# API_KEYS=key1,key2,key3
# Users connect:
claude mcp add --transport http your-server \
https://your-domain.com/mcp \
--header "Authorization: Bearer user_api_key"
Self-Host (GitHub):
# Share repo - users run their own instance
git clone your-repo
pnpm install && pnpm build
claude mcp add demo-server node -- $(pwd)/dist/servers/bare-metal/index.js
Implementation Details
Bare-Metal Approach
This implementation uses the official MCP SDK minimally, implementing core patterns manually:
Advantages:
- Full control and transparency
- Clear understanding of MCP mechanics
- Easy to customize for specific needs
- Minimal dependencies
Components:
- Manual request routing and handler registration
- Custom middleware pipeline
- Direct transport setup (stdio readline, Express HTTP)
- Explicit error handling patterns
Lines of Code: ~2500 (excluding tests)
Security Considerations
Production Checklist:
- [x] Input validation with Zod schemas
- [x] Path traversal protection for file access
- [x] Expression sanitization to prevent code injection
- [x] API key authentication (HTTP mode)
- [x] Rate limiting (token bucket algorithm)
- [x] Request tracing for audit logs
- [x] Secure environment variable handling
- [ ] HTTPS/TLS for HTTP transport (configure reverse proxy)
- [ ] API key rotation mechanism (add to config)
Extension Guide
Adding a new tool is straightforward:
1. Create handler (servers/bare-metal/handlers/MyTool.ts):
export class MyTool implements CapabilityHandler<MyParams, MyResult> {
validate(input: unknown): ValidationResult<MyParams> {
return validateWithSchema(MyParamsSchema, input);
}
async execute(params: MyParams, context: ExecutionContext): Promise<Result<MyResult>> {
// Your logic here
return { success: true, data: result };
}
handleError(error: Error): MCPError {
return { code: MCPErrorCode.InternalError, message: error.message };
}
}
2. Register in index.ts:
toolRegistry.register('my_tool', new MyTool(), {
name: 'my_tool',
description: 'Does something useful',
inputSchema: { /* JSON schema */ }
});
Done! Transport, routing, logging, and metrics work automatically.
Testing
# Run all tests
pnpm test
# Watch mode
pnpm test:watch
# Coverage report
pnpm test:coverage
Tests cover:
- Unit tests for all handlers
- Security validation (path traversal, injection attempts)
- Middleware functionality (auth, rate limiting)
- Integration tests for both transports
Project Structure
/mcp-demo
/shared # Common utilities
/types # TypeScript interfaces
/security # Validation, sanitization
/observability # Logging, metrics, tracing
/config # Configuration management
/utils # HTTP client, helpers
/servers/bare-metal # Bare-metal implementation
/handlers # Tool and resource handlers
/transport # Stdio and HTTP transports
/middleware # Auth, rate limiting
/registry # Capability registration
/core # Message routing
index.ts # Main entry point
/tests # Test suites
/docs # Specifications
License
MIT
Contributing
Contributions welcome! This project is designed as a learning resource and template for production MCP servers.
Focus areas:
- Additional capability examples (database, caching, etc.)
- Enhanced security patterns
- Performance optimizations
- Framework comparison implementations
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.