mcp_server_developer（MCP server 开发者）

MCP Server Developer

Expert implementation of Model Context Protocol (MCP) servers for Claude Desktop integration. This skill provides comprehensive guidance for building production-ready MCP servers with TypeScript, including architecture design, endpoint implementation, database integration, telemetry tracking, and distribution via npx.

Core Competencies

1. MCP Server Architecture

• Scaffold Generation: TypeScript project setup with proper MCP protocol structure
• Server Configuration: MCP protocol compliance, JSON-RPC handling, stdio transport
• Endpoint Design: RESTful-style resource and tool endpoints following MCP spec
• Type Safety: Strict TypeScript types for requests, responses, and schemas
• Error Handling: Comprehensive error catching with proper MCP error responses

2. Database Integration

• D1 SQLite Setup: Schema design optimized for edge deployment
• Query Optimization: <500ms latency targets for all database operations
• FTS5 Search: Full-text search implementation for model discovery
• Data Modeling: Efficient table structures with proper indexing
• Migration Scripts: Version-controlled schema evolution

3. Telemetry & Analytics

• Event Tracking: Structured logging for install, usage, and performance metrics
• Latency Monitoring: Request timing with percentile tracking (p50, p95, p99)
• User Analytics: Privacy-respecting usage patterns without PII
• Error Reporting: Structured error logging for debugging
• Success Metrics: WAU (Weekly Active Users) and engagement tracking

4. NPX Distribution

• Package Configuration: package.json with proper bin entries for npx execution
• Version Management: Semantic versioning with changelog automation
• Publishing Workflow: npm registry deployment with CI/CD integration
• Update Strategy: Non-breaking updates with deprecation notices
• Installation Testing: Cross-platform verification (macOS, Windows, Linux)

Implementation Guidelines

Phase 0: Server Scaffold (Week 1)

1. Initialize TypeScript Project npm init -y npm install --save-dev typescript @types/node npx tsc --init
2. MCP Server Setup

- Install MCP SDK: npm install @modelcontextprotocol/sdk - Create server entry point (src/index.ts) - Implement stdio transport handler - Add basic resource/tool endpoints

1. Type Definitions interface MCPServer {name: string; version: string; resources: Resource[]; tools: Tool[];} interface Resource {uri: string; name: string; description: string; mimeType?: string;} interface Tool {name: string; description: string; inputSchema: JSONSchema;}

Phase 1: Core Functionality (Week 2)

1. Database Schema CREATE TABLE mental_models (id TEXT PRIMARY KEY, code TEXT UNIQUE NOT NULL, transformation_class TEXT NOT NULL, name TEXT NOT NULL, description TEXT, use_cases JSON, difficulty_tier INTEGER, prerequisites JSON, created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP); CREATE VIRTUAL TABLE models_fts USING fts5(code, name, description, use_cases);
2. Endpoint Implementation

- GET resource endpoints (models list, transformations) - POST tool endpoints (analyze, decompose, synthesize) - Error handling with MCP-compliant responses - Input validation with Zod or similar

1. Transformation Logic

- Perspective (P): Multi-viewpoint analysis - Inversion (IN): Reverse assumption mapping - Composition (CO): Integration synthesis - Decomposition (DE): Component breakdown - Recursion (RE): Feedback loop identification - Meta-Systems (SY): Systems-of-systems coordination

Phase 2: Documentation & Examples (Week 3)

1. Quick Start Guide (5-minute setup) ``` # Quick Start 1. Install via npx: `bash npx @hummbl/mcp-server `` `

1. Configure Claude Desktop: Add to claude_desktop_config.json: {"mcpServers": {"hummbl": {"command": "npx", "args": ["@hummbl/mcp-server"]}}} 2. Test in Claude: "Use the perspective transformation on this problem: [your problem]"

1. Runnable Examples

- Example 1: Code analysis with decomposition - Example 2: Strategy synthesis with composition - Example 3: Decision making with perspective + inversion

1. Troubleshooting Guide

- Installation failures (permissions, npm config) - Claude Desktop connection issues - Latency/performance problems - Error message reference

Phase 3: Distribution & Testing (Week 4)

1. NPX Setup {"name": "@hummbl/mcp-server", "version": "1.0.0", "bin": {"hummbl-mcp": "./dist/index.js"}, "files": ["dist", "README.md", "LICENSE"], "publishConfig": {"access": "public"}}
2. CI/CD Pipeline

- GitHub Actions for automated testing - Automated npm publishing on release tags - Cross-platform binary testing - Documentation deployment

1. Beta Testing

- 3-5 beta testers minimum - Structured feedback collection - Performance benchmarking - Edge case identification

Quality Gates

Functional Requirements ✅

• All MCP endpoints respond correctly to spec
• Database queries complete <500ms (p95)
• Zero crashes during 24-hour stability test
• Error handling covers all edge cases
• Type safety: 100% TypeScript strict mode

Documentation Requirements ✅

• Quick start achieves first run in <5 minutes
• All examples run without modification
• Troubleshooting covers 90%+ of common issues
• API reference is complete and accurate
• Inline code comments for complex logic

Adoption Requirements ✅

• 10+ Weekly Active Users (WAU) by target date
• 3+ beta testers provide positive feedback
• Zero P0 bugs in production after Week 2
• Average user rating ≥4.0/5.0
• MCP directory listing approved

MCP Protocol Compliance

Required Capabilities

1. Resources: Read-only data access (models, transformations)
2. Tools: Interactive operations (analyze, decompose, synthesize)
3. Prompts: Pre-defined workflows (optional in Phase 0)

Transport Layer

• stdio: Standard input/output for Claude Desktop
• JSON-RPC 2.0: All requests/responses follow spec
• Error Codes: Proper HTTP-style error codes

Security Best Practices

• Input sanitization for all user-provided data
• Rate limiting to prevent abuse
• No external network calls without explicit permission
• Minimal permissions model
• Audit logging for all operations

Performance Optimization

Latency Targets

• p50: <100ms for all endpoints
• p95: <500ms for database queries
• p99: <1000ms for complex transformations

Optimization Strategies

1. Database: Index all frequently queried columns
2. Caching: In-memory LRU cache for hot data
3. Lazy Loading: Load resources only when needed
4. Batch Operations: Group similar queries
5. Connection Pooling: Reuse database connections

Telemetry Schema

interface TelemetryEvent {
  event_name: 'mcp_install' | 'mcp_success_run' | 'api_call' | 'doc_view';
  user_id: string; // Anonymized hash
  ts: string; // ISO 8601
  meta: {
    client?: string; // Claude Desktop version
    version?: string; // MCP server version
    model?: string; // e.g., "CO4"
    lat_ms?: number; // Latency in milliseconds
    endpoint?: string; // Endpoint called
    status?: number; // HTTP-style status code
  };
}

Distribution Checklist

Pre-Publication ✅

• Version number updated (semantic versioning)
• CHANGELOG.md updated with release notes
• README.md reviewed and accurate
• LICENSE file present (MIT recommended)
• package.json metadata complete
• Dependencies audit clean (npm audit)
• Bundle size acceptable (<5MB)

MCP Directory Submission ✅

• Clear description (160 characters max)
• Category selection appropriate
• Screenshots/demos prepared
• Usage examples documented
• Support contact provided

Post-Publication ✅

• NPM package downloadable via npx
• GitHub release created with notes
• Documentation site updated
• Community announcement posted
• Beta testers notified

Common Pitfalls & Solutions

Issue: Claude Desktop doesn't detect server

Solution: Verify stdio transport is properly initialized and JSON-RPC responses are formatted correctly.

Issue: High latency (>1s)

Solution: Add database indexes, implement caching, or optimize query complexity.

Issue: Installation fails on Windows

Solution: Test with cross-platform paths, avoid shell-specific commands, use cross-env for environment variables.

Issue: Breaking changes between versions

Solution: Follow semantic versioning strictly, maintain backwards compatibility, provide migration guides.

Examples

Example 1: Minimal MCP Server

import { Server } from '@modelcontextprotocol/sdk/server/index.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';

const server = new Server({
  name: 'hummbl-mcp-server',
  version: '1.0.0',
}, {
  capabilities: {
    resources: {},
    tools: {},
  },
});

// Register resource: List mental models
server.setRequestHandler('resources/list', async () => ({
  resources: [
    {
      uri: 'hummbl://models',
      name: 'Mental Models',
      description: 'BASE120 mental model collection',
      mimeType: 'application/json',
    },
  ],
}));

// Register tool: Analyze with perspective transformation
server.setRequestHandler('tools/call', async (request) => {
  if (request.params.name === 'analyze-perspective') {
    const { text } = request.params.arguments;
    // Transformation logic here
    return {
      content: [{
        type: 'text',
        text: `Perspective analysis: ${text}`,
      }],
    };
  }
});

// Start server
const transport = new StdioServerTransport();
await server.connect(transport);

Example 2: Database Integration

import Database from 'better-sqlite3';

const db = new Database('hummbl.db');

// Initialize schema
db.exec(`
  CREATE TABLE IF NOT EXISTS mental_models (
    id TEXT PRIMARY KEY,
    code TEXT UNIQUE NOT NULL,
    transformation_class TEXT NOT NULL,
    name TEXT NOT NULL,
    description TEXT
  );

  CREATE INDEX IF NOT EXISTS idx_transformation
  ON mental_models(transformation_class);
`);

// Query with prepared statement
const getModelsByTransformation = db.prepare(`
  SELECT * FROM mental_models
  WHERE transformation_class = ?
  ORDER BY code
`);

const perspectiveModels = getModelsByTransformation.all('P');

Example 3: Telemetry Logging

function logTelemetry(event: TelemetryEvent) {
  const logEntry = {
    ...event,
    ts: new Date().toISOString(),
  };

  // Log to console (development)
  if (process.env.NODE_ENV === 'development') {
    console.log(JSON.stringify(logEntry));
  }

  // Send to analytics service (production)
  if (process.env.NODE_ENV === 'production') {
    // TODO: Send to analytics endpoint
  }
}

// Usage
logTelemetry({
  event_name: 'api_call',
  user_id: hashUserId(userId),
  ts: new Date().toISOString(),
  meta: {
    endpoint: '/tools/analyze-perspective',
    lat_ms: 245,
    status: 200,
  },
});

Resources

• MCP Specification: https://spec.modelcontextprotocol.io/
• MCP SDK: https://github.com/modelcontextprotocol/typescript-sdk
• Claude Desktop Config: https://docs.claude.com/en/docs/agents-and-tools/agent-skills
• TypeScript Best Practices: https://typescript-lang.org/docs/handbook/intro.html
• NPM Publishing Guide: https://docs.npmjs.com/packages-and-modules/contributing-packages-to-the-registry

Success Criteria

Phase 0 is successful when:

1. ✅ npx @hummbl/mcp-server runs cleanly on any machine
2. ✅ 10+ developers using it weekly by target date
3. ✅ 3 compelling examples work out-of-the-box
4. ✅ Zero P0 bugs after Week 2
5. ✅ Clear path to Phase 1 identified

Phase 0 fails if:

1. ❌ <10 users after 60 days → Pivot to consulting focus
2. ❌ >5 P0 bugs in first 30 days → Architecture review needed *MCP Server Developer v1.1 - Enhanced*

🔄 Workflow

Kaynak: Anthropic MCP SDK Best Practices

Aşama 1: Architecture & Setup

• Scaffolding: npm init ve TypeScript config ayarlarını yap.
• Types: Resource ve Tool tiplerini strict mode ile tanımla.
• Transport: Stdio transport'u onerror handler ile güvenli hale getir.

Aşama 2: Development Loop

• Hot Reload: Geliştirme sırasında nodemon veya watch modunu kullan.
• Inspector: npx @modelcontextprotocol/inspector ile canlı debug yap.
• Logging: Structured logging (JSON) ekle ama stdio'yu kirletme (stderr kullan).

Aşama 3: Release

• Distribution: Paketi npm'e veya npx ile çalışacak şekilde publish et.
• Docs: README.md'ye claude_desktop_config.json örneğini ekle.
• Analytics: Anonim kullanım verisi toplama (opt-in) mekanizmasını kur.

Kontrol Noktaları