Building Custom MCP Servers: Extend Cursor with Your Tools

Building Custom MCP Servers: Extend Cursor with Your Tools

Model Context Protocol (MCP) servers are the bridge between Cursor and your external tools. While Cursor comes with many pre-built MCP servers, the real power lies in creating custom servers tailored to your organization’s unique needs.

Understanding MCP Architecture

MCP follows a client-server model where:

• Cursor acts as the client, sending requests for tools and data
• Your MCP server responds with capabilities and executes actions
• Communication happens via stdio, SSE, or HTTP transports

Note

Key Insight: MCP servers can be written in any language that can handle JSON and communicate via stdout or HTTP. Python, TypeScript, and Go are popular choices.

Your First Custom MCP Server

Let’s build a simple MCP server that integrates with a hypothetical internal API.

Step 1: Basic Server Structure (TypeScript)

TypeScript

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

// Create server instance
const server = new McpServer({
  name: "internal-api-server",
  version: "1.0.0",
  description: "Connects Cursor to our internal systems"
});

// Define a tool
server.tool(
  "get_user_info",
  "Fetch user information from internal API",
  {
    userId: z.string().describe("The user's ID")
  },
  async ({ userId }) => {
    // Your API call here
    const response = await fetch(`https://api.internal.com/users/${userId}`);
    const data = await response.json();

    return {
      content: [{
        type: "text",
        text: JSON.stringify(data, null, 2)
      }]
    };
  }
);

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

Python

from mcp.server import Server
from mcp.server.stdio import StdioServerTransport
from mcp.tools import Tool
import asyncio
import json
import aiohttp

# Create server instance
server = Server(
    name="internal-api-server",
    version="1.0.0"
)

# Define a tool
@server.tool()
async def get_user_info(userId: str) -> str:
    """Fetch user information from internal API"""
    async with aiohttp.ClientSession() as session:
        async with session.get(f"https://api.internal.com/users/{userId}") as response:
            data = await response.json()
            return json.dumps(data, indent=2)

# Start the server
async def main():
    transport = StdioServerTransport()
    await server.connect(transport)
    await transport.serve()

if __name__ == "__main__":
    asyncio.run(main())

Step 2: Configure in Cursor

Add to your ~/.cursor/mcp.json:

{
  "mcpServers": {
    "internal-api": {
      "command": "node",
      "args": ["/path/to/your/server.js"],
      "env": {
        "API_KEY": "your-api-key-here"
      }
    }
  }
}

Advanced MCP Server Patterns

Pattern 1: Multi-Tool Server

Create servers that expose multiple related tools:

// Database analysis MCP server
const dbServer = new McpServer({
  name: "database-analyzer",
  version: "1.0.0"
});

// Tool 1: Schema inspector
dbServer.tool(
  "inspect_schema",
  "Get database schema information",
  { table: z.string() },
  async ({ table }) => {
    const schema = await db.getTableSchema(table);
    return formatSchemaResponse(schema);
  }
);

// Tool 2: Query analyzer
dbServer.tool(
  "analyze_query",
  "Analyze SQL query performance",
  { query: z.string() },
  async ({ query }) => {
    const plan = await db.explainQuery(query);
    return formatQueryPlan(plan);
  }
);

// Tool 3: Data profiler
dbServer.tool(
  "profile_data",
  "Profile data distribution in a table",
  {
    table: z.string(),
    column: z.string().optional()
  },
  async ({ table, column }) => {
    const profile = await db.profileData(table, column);
    return formatDataProfile(profile);
  }
);

Pattern 2: Stateful MCP Server

Build servers that maintain state across requests:

class StatefulMcpServer {
  private sessions: Map<string, SessionData> = new Map();
  private server: McpServer;

  constructor() {
    this.server = new McpServer({
      name: "stateful-workflow",
      version: "1.0.0"
    });

    this.setupTools();
  }

  private setupTools() {
    // Start a session
    this.server.tool(
      "start_session",
      "Initialize a new workflow session",
      { workflowType: z.string() },
      async ({ workflowType }) => {
        const sessionId = generateId();
        this.sessions.set(sessionId, {
          type: workflowType,
          state: 'initialized',
          data: {}
        });

        return {
          content: [{
            type: "text",
            text: `Session started: ${sessionId}`
          }]
        };
      }
    );

    // Execute workflow steps
    this.server.tool(
      "execute_step",
      "Execute next step in workflow",
      {
        sessionId: z.string(),
        input: z.any()
      },
      async ({ sessionId, input }) => {
        const session = this.sessions.get(sessionId);
        if (!session) throw new Error("Session not found");

        // Process based on current state
        const result = await this.processWorkflowStep(session, input);

        return {
          content: [{
            type: "text",
            text: JSON.stringify(result)
          }]
        };
      }
    );
  }
}

Pattern 3: Resource Provider Server

Expose dynamic resources that Cursor can browse:

// Documentation server that provides browseable resources
server.resource(
  "api_docs",
  "Browse API documentation",
  async (uri) => {
    // Parse URI to determine what doc to show
    const path = uri.replace("api-docs://", "");

    if (path === "/") {
      // Return list of available APIs
      return {
        content: [{
          type: "text",
          text: await getApiIndex()
        }]
      };
    }

    // Return specific API documentation
    const doc = await getApiDoc(path);
    return {
      content: [{
        type: "text",
        text: doc
      }]
    };
  }
);

// Resources can be hierarchical
server.listResources = async () => {
  return [
    { uri: "api-docs://users", name: "User API" },
    { uri: "api-docs://payments", name: "Payment API" },
    { uri: "api-docs://analytics", name: "Analytics API" }
  ];
};

Real-World Custom MCP Examples

Example 1: Security Scanner MCP

A security team built an MCP server for automated security analysis:

• Directorysecurity-scanner-mcp/

  • Directorysrc/

    • index.ts
    • Directoryscanners/

      • dependencies.ts
      • secrets.ts
      • vulnerabilities.ts
    • Directoryreporters/

      • markdown.ts
      • json.ts
  • package.json
  • mcp.json

// Main security scanner implementation
class SecurityScannerMcp {
  private scanners = [
    new DependencyScanner(),
    new SecretScanner(),
    new VulnerabilityScanner()
  ];

  async setupTools() {
    // Comprehensive security scan
    this.server.tool(
      "security_scan",
      "Run comprehensive security analysis",
      {
        directory: z.string(),
        scanTypes: z.array(z.enum(['deps', 'secrets', 'vulns'])).optional()
      },
      async ({ directory, scanTypes }) => {
        const results = [];

        for (const scanner of this.scanners) {
          if (!scanTypes || scanTypes.includes(scanner.type)) {
            const findings = await scanner.scan(directory);
            results.push(...findings);
          }
        }

        // Generate report
        const report = new SecurityReport(results);
        return {
          content: [{
            type: "text",
            text: report.toMarkdown()
          }]
        };
      }
    );

    // Check specific file
    this.server.tool(
      "check_file_security",
      "Security check a specific file",
      {
        filePath: z.string()
      },
      async ({ filePath }) => {
        const issues = await this.quickScan(filePath);
        return formatIssues(issues);
      }
    );
  }
}

Example 2: Performance Monitoring MCP

Server Code

// Performance monitoring MCP that connects to APM tools
class PerformanceMcp {
  private apmClient: APMClient;

  constructor() {
    this.apmClient = new APMClient({
      endpoint: process.env.APM_ENDPOINT,
      apiKey: process.env.APM_API_KEY
    });
  }

  setupTools() {
    // Get performance metrics
    this.server.tool(
      "get_performance_metrics",
      "Fetch current performance metrics",
      {
        service: z.string(),
        timeRange: z.enum(['1h', '24h', '7d']).default('1h')
      },
      async ({ service, timeRange }) => {
        const metrics = await this.apmClient.getMetrics({
          service,
          from: getTimeFromRange(timeRange),
          to: new Date()
        });

        return {
          content: [{
            type: "text",
            text: this.formatMetrics(metrics)
          }]
        };
      }
    );

    // Analyze slow endpoints
    this.server.tool(
      "analyze_slow_endpoints",
      "Find and analyze slow API endpoints",
      {
        threshold: z.number().default(1000),
        limit: z.number().default(10)
      },
      async ({ threshold, limit }) => {
        const slowEndpoints = await this.apmClient.getSlowTransactions({
          threshold,
          limit
        });

        const analysis = await this.analyzeEndpoints(slowEndpoints);
        return {
          content: [{
            type: "text",
            text: analysis
          }]
        };
      }
    );
  }
}

Usage in Cursor

// In Cursor chat:
"Check performance metrics for the payment service"

// MCP server responds with:
/*
Payment Service Performance (Last 1h):
- Response Time: p50=145ms, p95=623ms, p99=1823ms
- Throughput: 1,234 req/min
- Error Rate: 0.23%
- CPU Usage: 67%
- Memory: 2.3GB / 4GB

Top Slow Endpoints:
1. POST /api/payment/process - 1823ms (p99)
2. GET /api/payment/history - 967ms (p99)
*/

"Analyze why the payment processing endpoint is slow"

// Detailed analysis with traces and suggestions

Example 3: Custom Deploy MCP

// Deployment automation MCP
class DeploymentMcp {
  setupTools() {
    // Deploy to environment
    this.server.tool(
      "deploy",
      "Deploy application to specified environment",
      {
        environment: z.enum(['dev', 'staging', 'prod']),
        version: z.string().optional(),
        dryRun: z.boolean().default(false)
      },
      async ({ environment, version, dryRun }) => {
        // Pre-deployment checks
        const checks = await this.runPreDeployChecks(environment);
        if (!checks.passed) {
          return {
            content: [{
              type: "text",
              text: `Pre-deploy checks failed:\n${checks.failures.join('\n')}`
            }]
          };
        }

        // Execute deployment
        if (!dryRun) {
          const result = await this.deploy(environment, version);
          await this.notifyTeam(environment, result);

          return {
            content: [{
              type: "text",
              text: this.formatDeployResult(result)
            }]
          };
        }

        return {
          content: [{
            type: "text",
            text: "Dry run completed successfully"
          }]
        };
      }
    );
  }
}

Security Best Practices

1. Authentication & Authorization

// Implement auth in your MCP server
class SecureMcpServer {
  private async authenticate(headers: Headers): Promise<boolean> {
    const token = headers.get('Authorization')?.replace('Bearer ', '');
    if (!token) return false;

    // Verify token with your auth service
    return await this.authService.verifyToken(token);
  }

  async handleRequest(request: Request) {
    if (!await this.authenticate(request.headers)) {
      throw new Error('Unauthorized');
    }

    // Process authenticated request
  }
}

2. Input Validation

// Always validate and sanitize inputs
server.tool(
  "execute_query",
  "Run a database query",
  {
    query: z.string()
      .min(1)
      .max(1000)
      .refine(
        (q) => !q.match(/DROP|DELETE|TRUNCATE/i),
        "Destructive operations not allowed"
      )
  },
  async ({ query }) => {
    // Additional validation
    const sanitized = sanitizeSQL(query);
    const result = await db.query(sanitized);
    return formatResult(result);
  }
);

3. Secret Management

// Use environment variables for secrets
const config = {
  apiKey: process.env.API_KEY,
  dbPassword: process.env.DB_PASSWORD,
  // Never hardcode secrets!
};

// Validate required secrets on startup
function validateEnvironment() {
  const required = ['API_KEY', 'DB_PASSWORD'];
  const missing = required.filter(key => !process.env[key]);

  if (missing.length > 0) {
    throw new Error(`Missing required env vars: ${missing.join(', ')}`);
  }
}

Testing Your MCP Server

Unit Testing Tools

import { MockMcpClient } from '@mcp/testing';

describe('SecurityScannerMcp', () => {
  let server: SecurityScannerMcp;
  let client: MockMcpClient;

  beforeEach(() => {
    server = new SecurityScannerMcp();
    client = new MockMcpClient(server);
  });

  test('security_scan finds vulnerabilities', async () => {
    const result = await client.callTool('security_scan', {
      directory: './test-project',
      scanTypes: ['vulns']
    });

    expect(result.content[0].text).toContain('Found 3 vulnerabilities');
  });
});

Integration Testing

// Test with actual Cursor connection
async function testWithCursor() {
  // Start your MCP server
  const serverProcess = spawn('node', ['./dist/index.js']);

  // Configure test Cursor instance
  const testConfig = {
    mcpServers: {
      "test-server": {
        command: "node",
        args: ["./dist/index.js"]
      }
    }
  };

  // Run test scenarios
  // ...
}

Deployment Strategies

Local Development

# Development setup
npm run build:watch

# Configure your local MCP server in Cursor settings
# Settings > Features > MCP > Add Server
{
  "mcpServers": {
    "my-server-dev": {
      "command": "node",
      "args": ["./dist/index.js"],
      "env": {
        "NODE_ENV": "development",
        "DEBUG": "mcp:*"
      }
    }
  }
}

Team Distribution

NPM Package

package.json

{
  "name": "@company/mcp-internal-tools",
  "version": "1.0.0",
  "bin": {
    "mcp-internal": "./dist/index.js"
  },
  "publishConfig": {
    "registry": "https://npm.company.com"
  }
}

// Team installs globally
npm install -g @company/mcp-internal-tools

// Configure in Cursor
{
  "mcpServers": {
    "internal-tools": {
      "command": "mcp-internal"
    }
  }
}

Remote Server

// Deploy as HTTP server
const server = new McpHttpServer({
  port: 3000,
  auth: new OAuthProvider()
});

// Team configures remote URL
{
  "mcpServers": {
    "internal-tools": {
      "transport": "sse",
      "url": "https://mcp.company.com/sse"
    }
  }
}

Performance Optimization

1. Response Streaming

For large responses, use streaming:

server.tool(
  "analyze_large_dataset",
  "Analyze a large dataset",
  { datasetId: z.string() },
  async function* ({ datasetId }) {
    const dataset = await loadDataset(datasetId);

    // Stream results as they're processed
    for await (const batch of processBatches(dataset)) {
      yield {
        content: [{
          type: "text",
          text: `Processed batch: ${JSON.stringify(batch.summary)}\n`
        }]
      };
    }
  }
);

2. Caching Strategies

class CachedMcpServer {
  private cache = new LRUCache<string, any>({
    max: 100,
    ttl: 1000 * 60 * 5 // 5 minutes
  });

  async handleExpensiveOperation(params: any) {
    const cacheKey = JSON.stringify(params);

    // Check cache first
    const cached = this.cache.get(cacheKey);
    if (cached) return cached;

    // Perform expensive operation
    const result = await this.expensiveOperation(params);

    // Cache result
    this.cache.set(cacheKey, result);
    return result;
  }
}

Debugging MCP Servers

Enable Debug Logging

// Use debug package for detailed logging
import debug from 'debug';
const log = debug('mcp:my-server');

server.on('tool_call', (tool, params) => {
  log('Tool called:', tool, params);
});

// Run with DEBUG=mcp:* to see logs

MCP Inspector Tool

# Install MCP inspector
npm install -g @mcp/inspector

# Test your server
mcp-inspect ./my-server.js

# Interactive testing interface
> list_tools
> call_tool get_user_info {"userId": "123"}

Common Pitfalls and Solutions

Timeout Issues

Problem: Long-running operations timeout Solution: Implement progress reporting or break into smaller operations

Memory Leaks

Problem: Server memory usage grows over time Solution: Properly cleanup resources, use WeakMaps for caches

Error Handling

Problem: Errors crash the server Solution: Wrap all tool handlers in try-catch, return friendly errors

Next Steps

Now that you can build custom MCP servers:

1. Identify Integration Opportunities: List your team’s most-used tools
2. Start Simple: Build a read-only MCP server first
3. Iterate Based on Usage: Add features as needed
4. Share with Team: Package and distribute to multiply impact

Remember: The best MCP servers solve real problems. Start with your biggest pain point and build from there. Each custom integration makes Cursor more powerful for your specific workflow.