MCP (Model Context Protocol) - AI-Native Server Development

Overview

Model Context Protocol (MCP) is an open standard for connecting AI assistants to external data sources and tools. Build servers that expose tools (functions LLMs can call), resources (data LLMs can read), and prompts (templates LLMs can use).

Key Concepts:

• Tools: Functions LLMs can execute (read files, query APIs, run commands)
• Resources: Data sources LLMs can access (files, databases, APIs)
• Prompts: Reusable templates with arguments for common tasks
• Client-Server: MCP servers expose capabilities, clients (like Claude Desktop) consume them
• Transport: STDIO (local), SSE (Server-Sent Events), HTTP (network)

Official SDKs:

• TypeScript: @modelcontextprotocol/sdk
• Python: mcp

Installation:

# TypeScript server
npx @modelcontextprotocol/create-server@latest my-server
cd my-server && npm install

# Python server
pip install mcp
# Or use uv (recommended)
uv pip install mcp

Quick Start - TypeScript Server

1. Create Server with CLI

# Interactive setup
npx @modelcontextprotocol/create-server@latest my-filesystem-server

# Options prompt:
# - Server name: my-filesystem-server
# - Language: TypeScript
# - Include example tools: Yes

2. Define Tools

// src/index.ts
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import {
  CallToolRequestSchema,
  ListToolsRequestSchema,
} from "@modelcontextprotocol/sdk/types.js";
import * as fs from "fs/promises";
import * as path from "path";

const server = new Server(
  {
    name: "filesystem-server",
    version: "1.0.0",
  },
  {
    capabilities: {
      tools: {},
    },
  }
);

// List available tools
server.setRequestHandler(ListToolsRequestSchema, async () => {
  return {
    tools: [
      {
        name: "read_file",
        description: "Read contents of a file",
        inputSchema: {
          type: "object",
          properties: {
            path: {
              type: "string",
              description: "Path to the file to read",
            },
          },
          required: ["path"],
        },
      },
      {
        name: "list_directory",
        description: "List contents of a directory",
        inputSchema: {
          type: "object",
          properties: {
            path: {
              type: "string",
              description: "Directory path to list",
            },
          },
          required: ["path"],
        },
      },
    ],
  };
});

// Handle tool calls
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;

  switch (name) {
    case "read_file": {
      const filePath = args.path as string;
      const content = await fs.readFile(filePath, "utf-8");
      return {
        content: [{ type: "text", text: content }],
      };
    }

    case "list_directory": {
      const dirPath = args.path as string;
      const entries = await fs.readdir(dirPath, { withFileTypes: true });
      const listing = entries
        .map((entry) => `${entry.isDirectory() ? "📁" : "📄"} ${entry.name}`)
        .join("\n");
      return {
        content: [{ type: "text", text: listing }],
      };
    }

    default:
      throw new Error(`Unknown tool: ${name}`);
  }
});

// Start server
async function main() {
  const transport = new StdioServerTransport();
  await server.connect(transport);
  console.error("Filesystem MCP server running on stdio");
}

main();

3. Configure Claude Desktop

// ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
// %APPDATA%/Claude/claude_desktop_config.json (Windows)
{
  "mcpServers": {
    "filesystem": {
      "command": "node",
      "args": ["/absolute/path/to/my-filesystem-server/build/index.js"]
    }
  }
}

4. Build and Test

# Build TypeScript
npm run build

# Restart Claude Desktop (Cmd+Q and reopen)
# Server appears in 🔌 menu

Quick Start - Python Server

1. Create Server

# server.py
import asyncio
from mcp.server import Server
from mcp.server.stdio import stdio_server
from mcp.types import Tool, TextContent
import json
import os

# Create server instance
app = Server("filesystem-server")

# Define tools
@app.list_tools()
async def list_tools() -> list[Tool]:
    return [
        Tool(
            name="read_file",
            description="Read contents of a file",
            inputSchema={
                "type": "object",
                "properties": {
                    "path": {"type": "string", "description": "File path to read"}
                },
                "required": ["path"],
            },
        ),
        Tool(
            name="list_directory",
            description="List directory contents",
            inputSchema={
                "type": "object",
                "properties": {
                    "path": {"type": "string", "description": "Directory path"}
                },
                "required": ["path"],
            },
        ),
    ]

# Handle tool calls
@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    if name == "read_file":
        file_path = arguments["path"]
        with open(file_path, "r") as f:
            content = f.read()
        return [TextContent(type="text", text=content)]

    elif name == "list_directory":
        dir_path = arguments["path"]
        entries = os.listdir(dir_path)
        listing = "\n".join(
            f"{'📁' if os.path.isdir(os.path.join(dir_path, e)) else '📄'} {e}"
            for e in entries
        )
        return [TextContent(type="text", text=listing)]

    else:
        raise ValueError(f"Unknown tool: {name}")

# Start server
async def main():
    async with stdio_server() as (read_stream, write_stream):
        await app.run(read_stream, write_stream, app.create_initialization_options())

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

2. Configure Claude Desktop

{
  "mcpServers": {
    "filesystem": {
      "command": "python",
      "args": ["/absolute/path/to/server.py"]
    }
  }
}

3. Test

# Test server standalone
python server.py

# Restart Claude Desktop
# Tools appear in 🔌 menu

Resources - Exposing Data to LLMs

Resources provide read-only access to data sources.

TypeScript Resource Example

import {
  ListResourcesRequestSchema,
  ReadResourceRequestSchema,
} from "@modelcontextprotocol/sdk/types.js";

// List available resources
server.setRequestHandler(ListResourcesRequestSchema, async () => {
  return {
    resources: [
      {
        uri: "file:///docs/readme.md",
        name: "README",
        description: "Project README documentation",
        mimeType: "text/markdown",
      },
      {
        uri: "file:///config/settings.json",
        name: "Settings",
        description: "Application settings",
        mimeType: "application/json",
      },
    ],
  };
});

// Handle resource reads
server.setRequestHandler(ReadResourceRequestSchema, async (request) => {
  const uri = request.params.uri;

  if (uri.startsWith("file://")) {
    const filePath = uri.replace("file://", "");
    const content = await fs.readFile(filePath, "utf-8");
    return {
      contents: [
        {
          uri,
          mimeType: "text/plain",
          text: content,
        },
      ],
    };
  }

  throw new Error(`Unknown resource: ${uri}`);
});

Python Resource Example

from mcp.types import Resource, ResourceContent

@app.list_resources()
async def list_resources() -> list[Resource]:
    return [
        Resource(
            uri="file:///docs/readme.md",
            name="README",
            description="Project README",
            mimeType="text/markdown",
        ),
        Resource(
            uri="file:///config/settings.json",
            name="Settings",
            description="App settings",
            mimeType="application/json",
        ),
    ]

@app.read_resource()
async def read_resource(uri: str) -> ResourceContent:
    if uri.startswith("file://"):
        file_path = uri.replace("file://", "")
        with open(file_path, "r") as f:
            content = f.read()
        return ResourceContent(uri=uri, mimeType="text/plain", text=content)

    raise ValueError(f"Unknown resource: {uri}")

Prompts - Reusable Templates

Prompts are templates that LLMs can use with arguments.

TypeScript Prompt Example

import {
  ListPromptsRequestSchema,
  GetPromptRequestSchema,
} from "@modelcontextprotocol/sdk/types.js";

// List prompts
server.setRequestHandler(ListPromptsRequestSchema, async () => {
  return {
    prompts: [
      {
        name: "code_review",
        description: "Review code for best practices",
        arguments: [
          {
            name: "language",
            description: "Programming language",
            required: true,
          },
          {
            name: "code",
            description: "Code to review",
            required: true,
          },
        ],
      },
    ],
  };
});

// Handle prompt requests
server.setRequestHandler(GetPromptRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;

  if (name === "code_review") {
    const language = args?.language || "unknown";
    const code = args?.code || "";

    return {
      messages: [
        {
          role: "user",
          content: {
            type: "text",
            text: `Review this ${language} code for best practices, security issues, and improvements:\n\n\`\`\`${language}\n${code}\n\`\`\``,
          },
        },
      ],
    };
  }

  throw new Error(`Unknown prompt: ${name}`);
});

Python Prompt Example

from mcp.types import Prompt, PromptMessage, PromptArgument

@app.list_prompts()
async def list_prompts() -> list[Prompt]:
    return [
        Prompt(
            name="code_review",
            description="Review code for best practices",
            arguments=[
                PromptArgument(
                    name="language", description="Programming language", required=True
                ),
                PromptArgument(name="code", description="Code to review", required=True),
            ],
        )
    ]

@app.get_prompt()
async def get_prompt(name: str, arguments: dict) -> list[PromptMessage]:
    if name == "code_review":
        language = arguments.get("language", "unknown")
        code = arguments.get("code", "")

        return [
            PromptMessage(
                role="user",
                content={
                    "type": "text",
                    "text": f"Review this {language} code:\n\n```{language}\n{code}\n```",
                },
            )
        ]

    raise ValueError(f"Unknown prompt: {name}")

Advanced Patterns

1. Error Handling

// TypeScript
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  try {
    const { name, arguments: args } = request.params;

    if (name === "risky_operation") {
      // Validate inputs
      if (!args.required_param) {
        throw new Error("Missing required parameter: required_param");
      }

      // Perform operation with proper error handling
      const result = await performRiskyOperation(args.required_param);

      return {
        content: [{ type: "text", text: JSON.stringify(result) }],
      };
    }
  } catch (error) {
    // Return error to LLM with helpful message
    return {
      content: [
        {
          type: "text",
          text: `Error: ${error instanceof Error ? error.message : String(error)}`,
        },
      ],
      isError: true,
    };
  }
});

# Python
@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    try:
        if name == "risky_operation":
            if "required_param" not in arguments:
                raise ValueError("Missing required parameter: required_param")

            result = await perform_risky_operation(arguments["required_param"])
            return [TextContent(type="text", text=json.dumps(result))]

    except Exception as e:
        return [TextContent(type="text", text=f"Error: {str(e)}", isError=True)]

2. Async Operations

// TypeScript - Async API calls
import axios from "axios";

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;

  if (name === "fetch_api_data") {
    const url = args.url as string;
    const response = await axios.get(url);

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

# Python - Async database queries
import aiosqlite

@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    if name == "query_database":
        query = arguments["query"]

        async with aiosqlite.connect("database.db") as db:
            async with db.execute(query) as cursor:
                rows = await cursor.fetchall()
                result = json.dumps(rows)

        return [TextContent(type="text", text=result)]

3. Environment Variables & Configuration

// TypeScript - Load config from environment
import dotenv from "dotenv";
dotenv.config();

const API_KEY = process.env.API_KEY;
const BASE_URL = process.env.BASE_URL || "https://api.example.com";

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  if (request.params.name === "api_call") {
    const response = await fetch(`${BASE_URL}/endpoint`, {
      headers: { Authorization: `Bearer ${API_KEY}` },
    });
    // ...
  }
});

# Python - Load config from environment
import os
from dotenv import load_dotenv

load_dotenv()

API_KEY = os.getenv("API_KEY")
BASE_URL = os.getenv("BASE_URL", "https://api.example.com")

@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    if name == "api_call":
        headers = {"Authorization": f"Bearer {API_KEY}"}
        async with aiohttp.ClientSession() as session:
            async with session.get(f"{BASE_URL}/endpoint", headers=headers) as resp:
                data = await resp.json()
                return [TextContent(type="text", text=json.dumps(data))]

4. Streaming Responses (Large Data)

// TypeScript - Stream large file contents
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  if (request.params.name === "read_large_file") {
    const filePath = request.params.arguments.path as string;
    const stream = fs.createReadStream(filePath, { encoding: "utf-8" });

    let content = "";
    for await (const chunk of stream) {
      content += chunk;
      // Could yield chunks incrementally in future MCP versions
    }

    return {
      content: [{ type: "text", text: content }],
    };
  }
});

5. Dynamic Tool Registration

// TypeScript - Register tools from config
interface ToolConfig {
  name: string;
  description: string;
  schema: object;
  handler: (args: any) => Promise<any>;
}

const toolRegistry = new Map<string, ToolConfig>();

function registerTool(config: ToolConfig) {
  toolRegistry.set(config.name, config);
}

// Register custom tools
registerTool({
  name: "custom_tool",
  description: "Dynamically registered tool",
  schema: {
    type: "object",
    properties: {
      input: { type: "string" },
    },
  },
  handler: async (args) => {
    return { result: `Processed: ${args.input}` };
  },
});

// List tools dynamically
server.setRequestHandler(ListToolsRequestSchema, async () => {
  return {
    tools: Array.from(toolRegistry.values()).map((tool) => ({
      name: tool.name,
      description: tool.description,
      inputSchema: tool.schema,
    })),
  };
});

// Call tools dynamically
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const tool = toolRegistry.get(request.params.name);
  if (!tool) {
    throw new Error(`Unknown tool: ${request.params.name}`);
  }

  const result = await tool.handler(request.params.arguments);
  return {
    content: [{ type: "text", text: JSON.stringify(result) }],
  };
});

Transport Types

1. STDIO (Local Execution)

Default for Claude Desktop integration. Server runs as subprocess.

// Claude Desktop config
{
  "mcpServers": {
    "my-server": {
      "command": "node",
      "args": ["/path/to/server/build/index.js"],
      "env": {
        "API_KEY": "your-api-key"
      }
    }
  }
}

2. SSE (Server-Sent Events)

For long-running servers with HTTP transport.

// TypeScript SSE server
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { SSEServerTransport } from "@modelcontextprotocol/sdk/server/sse.js";
import express from "express";

const app = express();
const server = new Server(/* ... */);

app.get("/sse", async (req, res) => {
  const transport = new SSEServerTransport("/messages", res);
  await server.connect(transport);
});

app.post("/messages", async (req, res) => {
  // Handle incoming messages
});

app.listen(3000, () => {
  console.log("MCP server listening on http://localhost:3000");
});

// Claude Desktop config for SSE
{
  "mcpServers": {
    "remote-server": {
      "url": "http://localhost:3000/sse"
    }
  }
}

3. HTTP (Streamable HTTP)

For REST-style MCP servers.

# Python HTTP server with FastAPI
from fastapi import FastAPI
from mcp.server.fastapi import create_fastapi_app

app = FastAPI()
mcp_app = Server("http-server")

# Define tools/resources/prompts as usual
# ...

# Mount MCP routes
app.mount("/mcp", create_fastapi_app(mcp_app))

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)

Debugging MCP Servers

1. Server Logs

// TypeScript - Console.error for logs (STDOUT is for MCP protocol)
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  console.error(`Tool called: ${request.params.name}`);
  console.error(`Arguments: ${JSON.stringify(request.params.arguments)}`);
  // ...
});

# Python - Use logging module
import logging

logging.basicConfig(level=logging.DEBUG, filename="mcp-server.log")
logger = logging.getLogger(__name__)

@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    logger.debug(f"Tool called: {name}")
    logger.debug(f"Arguments: {arguments}")
    # ...

2. MCP Inspector

# Install MCP Inspector (browser-based debugging tool)
npm install -g @modelcontextprotocol/inspector

# Run inspector with your server
mcp-inspector node /path/to/server/build/index.js

# Opens browser at http://localhost:6274
# - Test tools manually
# - See request/response payloads
# - Debug JSON schema validation

3. Claude Desktop Logs

# macOS
tail -f ~/Library/Logs/Claude/mcp*.log

# Windows
# Check %APPDATA%/Claude/logs/

4. Test Server Standalone

// TypeScript - Add test harness
if (process.argv.includes("--test")) {
  // Simulate tool call
  const testRequest = {
    params: {
      name: "read_file",
      arguments: { path: "./test.txt" },
    },
  };

  server
    .setRequestHandler(CallToolRequestSchema, async (request) => {
      // ... your handler
    })
    .then((handler) => handler(testRequest))
    .then((result) => console.log(JSON.stringify(result, null, 2)))
    .catch((error) => console.error(error));
} else {
  // Normal server startup
  main();
}

# Test without Claude Desktop
node build/index.js --test

Real-World Server Examples

1. GitHub Integration Server

// github-server/src/index.ts
import { Octokit } from "@octokit/rest";

const octokit = new Octokit({ auth: process.env.GITHUB_TOKEN });

server.setRequestHandler(ListToolsRequestSchema, async () => {
  return {
    tools: [
      {
        name: "list_repos",
        description: "List repositories for a user/org",
        inputSchema: {
          type: "object",
          properties: {
            owner: { type: "string", description: "Username or org name" },
          },
          required: ["owner"],
        },
      },
      {
        name: "create_issue",
        description: "Create a GitHub issue",
        inputSchema: {
          type: "object",
          properties: {
            owner: { type: "string" },
            repo: { type: "string" },
            title: { type: "string" },
            body: { type: "string" },
          },
          required: ["owner", "repo", "title"],
        },
      },
    ],
  };
});

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;

  if (name === "list_repos") {
    const { data } = await octokit.repos.listForUser({
      username: args.owner as string,
    });
    const repos = data.map((r) => ({
      name: r.name,
      description: r.description,
      stars: r.stargazers_count,
    }));
    return {
      content: [{ type: "text", text: JSON.stringify(repos, null, 2) }],
    };
  }

  if (name === "create_issue") {
    const { data } = await octokit.issues.create({
      owner: args.owner as string,
      repo: args.repo as string,
      title: args.title as string,
      body: args.body as string,
    });
    return {
      content: [
        { type: "text", text: `Issue created: ${data.html_url}` },
      ],
    };
  }
});

2. Database Query Server

# database-server/server.py
import asyncpg
import json
from mcp.server import Server
from mcp.types import Tool, TextContent

app = Server("database-server")

async def get_db_pool():
    return await asyncpg.create_pool(
        host="localhost",
        database="mydb",
        user="user",
        password="password"
    )

@app.list_tools()
async def list_tools() -> list[Tool]:
    return [
        Tool(
            name="query",
            description="Execute SQL query (SELECT only)",
            inputSchema={
                "type": "object",
                "properties": {
                    "sql": {"type": "string", "description": "SQL query"},
                },
                "required": ["sql"],
            },
        ),
        Tool(
            name="list_tables",
            description="List all tables in database",
            inputSchema={"type": "object", "properties": {}},
        ),
    ]

@app.call_tool()
async def call_tool(name: str, arguments: dict) -> list[TextContent]:
    pool = await get_db_pool()

    if name == "query":
        sql = arguments["sql"]
        # Security: Only allow SELECT
        if not sql.strip().upper().startswith("SELECT"):
            raise ValueError("Only SELECT queries allowed")

        async with pool.acquire() as conn:
            rows = await conn.fetch(sql)
            result = [dict(row) for row in rows]

        return [TextContent(type="text", text=json.dumps(result, indent=2))]

    elif name == "list_tables":
        async with pool.acquire() as conn:
            tables = await conn.fetch("""
                SELECT table_name FROM information_schema.tables
                WHERE table_schema = 'public'
            """)
            result = [row["table_name"] for row in tables]

        return [TextContent(type="text", text=json.dumps(result, indent=2))]

3. Web Scraper Server

// scraper-server/src/index.ts
import * as cheerio from "cheerio";
import axios from "axios";

server.setRequestHandler(ListToolsRequestSchema, async () => {
  return {
    tools: [
      {
        name: "scrape_url",
        description: "Scrape content from a URL",
        inputSchema: {
          type: "object",
          properties: {
            url: { type: "string", description: "URL to scrape" },
            selector: {
              type: "string",
              description: "CSS selector for content",
            },
          },
          required: ["url"],
        },
      },
    ],
  };
});

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  if (request.params.name === "scrape_url") {
    const url = request.params.arguments.url as string;
    const selector = request.params.arguments.selector as string;

    const response = await axios.get(url);
    const $ = cheerio.load(response.data);

    let content: string;
    if (selector) {
      content = $(selector).text();
    } else {
      content = $("body").text();
    }

    return {
      content: [
        {
          type: "text",
          text: content.trim(),
        },
      ],
    };
  }
});

Best Practices

1. Tool Design

✅ DO:

• Use clear, descriptive tool names (read_file not rf)
• Provide detailed descriptions for LLM understanding
• Define comprehensive JSON schemas with descriptions
• Return structured data (JSON) when possible
• Handle errors gracefully with helpful messages

❌ DON'T:

• Create overly broad tools (split complex operations)
• Return massive payloads (paginate large datasets)
• Use ambiguous parameter names
• Assume LLM knows your domain-specific terminology

2. Security

Critical Rules:

• Validate ALL inputs (type, range, format)
• Sanitize file paths (prevent directory traversal)
• Use allowlists for commands/operations
• Never expose sensitive credentials in responses
• Implement rate limiting for expensive operations
• Use read-only access by default

// Example: Path validation
function validatePath(inputPath: string): string {
  const normalized = path.normalize(inputPath);
  const allowed = path.resolve("/safe/directory");

  if (!normalized.startsWith(allowed)) {
    throw new Error("Path outside allowed directory");
  }

  return normalized;
}

3. Performance

• Cache expensive operations
• Stream large responses when possible
• Use pagination for list operations
• Set reasonable timeouts
• Implement request queuing for rate-limited APIs

4. Testing

# Use MCP Inspector for manual testing
mcp-inspector node build/index.js

# Unit test tool handlers
npm test

# Integration test with Claude Desktop
# 1. Add to config
# 2. Restart Claude
# 3. Test in conversation

5. Documentation

• Document all tools in code comments
• Provide example usage in descriptions
• Include error scenarios in documentation
• Maintain a CHANGELOG for server updates

Common Pitfalls

❌ Writing to STDOUT (breaks MCP protocol):

// WRONG
console.log("Debug message");  // STDOUT is for MCP protocol

// CORRECT
console.error("Debug message");  // STDERR for logs

❌ Not handling errors:

// WRONG - unhandled promise rejection
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const data = await riskyOperation();  // Can throw
  return { content: [{ type: "text", text: data }] };
});

// CORRECT
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  try {
    const data = await riskyOperation();
    return { content: [{ type: "text", text: data }] };
  } catch (error) {
    return {
      content: [{ type: "text", text: `Error: ${error.message}` }],
      isError: true,
    };
  }
});

❌ Blocking operations:

# WRONG - synchronous file read blocks event loop
@app.call_tool()
async def call_tool(name: str, arguments: dict):
    with open("large_file.txt", "r") as f:  # Blocks!
        content = f.read()
    return [TextContent(type="text", text=content)]

# CORRECT - async file I/O
@app.call_tool()
async def call_tool(name: str, arguments: dict):
    async with aiofiles.open("large_file.txt", "r") as f:
        content = await f.read()
    return [TextContent(type="text", text=content)]

❌ Missing required fields in schema:

// WRONG - missing "required" field
{
  name: "search",
  inputSchema: {
    type: "object",
    properties: {
      query: { type: "string" }
    }
    // Missing: required: ["query"]
  }
}

Resources

• Official Docs: https://modelcontextprotocol.io
• TypeScript SDK: https://github.com/modelcontextprotocol/typescript-sdk
• Python SDK: https://github.com/modelcontextprotocol/python-sdk
• Server Examples: https://github.com/modelcontextprotocol/servers
• MCP Inspector: https://github.com/modelcontextprotocol/inspector
• Specification: https://spec.modelcontextprotocol.io

Related Skills

When building MCP servers, consider these complementary skills:

• typescript-core: TypeScript type safety, tsconfig optimization, and advanced patterns
• asyncio: Python async patterns for MCP servers with async/await
• openrouter: Alternative LLM API integration for multi-model support

Quick TypeScript MCP Patterns (Inlined for Standalone Use)

// Type-safe MCP server with TypeScript
import { Server } from '@modelcontextprotocol/sdk/server/index.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
import {
  CallToolRequestSchema,
  ListToolsRequestSchema,
} from '@modelcontextprotocol/sdk/types.js';
import { z } from 'zod';

// Define tool schemas with runtime validation
const SearchSchema = z.object({
  query: z.string().min(1).max(500),
  limit: z.number().int().min(1).max(100).default(10),
});

type SearchArgs = z.infer<typeof SearchSchema>;

class TypeSafeMCPServer {
  private server: Server;

  constructor() {
    this.server = new Server(
      {
        name: 'typed-search-server',
        version: '1.0.0',
      },
      {
        capabilities: {
          tools: {},
        },
      }
    );

    this.setupToolHandlers();
  }

  private setupToolHandlers() {
    // List tools with full type inference
    this.server.setRequestHandler(ListToolsRequestSchema, async () => ({
      tools: [
        {
          name: 'search',
          description: 'Search with type-safe parameters',
          inputSchema: {
            type: 'object',
            properties: {
              query: { type: 'string', minLength: 1, maxLength: 500 },
              limit: { type: 'number', minimum: 1, maximum: 100, default: 10 },
            },
            required: ['query'],
          },
        },
      ],
    }));

    // Type-safe tool execution
    this.server.setRequestHandler(CallToolRequestSchema, async (request) => {
      if (request.params.name === 'search') {
        // Runtime validation with Zod
        const args = SearchSchema.parse(request.params.arguments);

        // Type-safe implementation
        const results = await this.performSearch(args);

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

      throw new Error(`Unknown tool: ${request.params.name}`);
    });
  }

  private async performSearch(args: SearchArgs): Promise<Array<{ title: string; url: string }>> {
    // Implementation with full type safety
    // args.query is string, args.limit is number
    return [];
  }

  async run() {
    const transport = new StdioServerTransport();
    await this.server.connect(transport);
  }
}

// Start server
const server = new TypeSafeMCPServer();
server.run();

Quick Python Async MCP Patterns (Inlined for Standalone Use)

# Async MCP server with Python
import asyncio
import logging
from typing import Any
from mcp.server import Server
from mcp.server.stdio import stdio_server
from mcp.types import Tool, TextContent
from pydantic import BaseModel, Field

# Type-safe argument models with Pydantic
class SearchArgs(BaseModel):
    query: str = Field(..., min_length=1, max_length=500)
    limit: int = Field(10, ge=1, le=100)

class AsyncMCPServer:
    def __init__(self):
        self.server = Server("async-search-server")
        self.setup_handlers()

    def setup_handlers(self):
        @self.server.list_tools()
        async def list_tools() -> list[Tool]:
            return [
                Tool(
                    name="search",
                    description="Search with async processing",
                    inputSchema={
                        "type": "object",
                        "properties": {
                            "query": {"type": "string", "minLength": 1, "maxLength": 500},
                            "limit": {"type": "number", "minimum": 1, "maximum": 100, "default": 10},
                        },
                        "required": ["query"],
                    },
                )
            ]

        @self.server.call_tool()
        async def call_tool(name: str, arguments: dict[str, Any]) -> list[TextContent]:
            if name == "search":
                # Validate with Pydantic
                args = SearchArgs(**arguments)

                # Async implementation
                results = await self.perform_search(args)

                return [
                    TextContent(
                        type="text",
                        text=str(results)
                    )
                ]

            raise ValueError(f"Unknown tool: {name}")

    async def perform_search(self, args: SearchArgs) -> list[dict[str, str]]:
        """Async search implementation"""
        # Simulate async I/O
        await asyncio.sleep(0.1)

        # Use validated args (args.query is str, args.limit is int)
        return [
            {"title": f"Result for {args.query}", "url": "https://example.com"}
        ]

async def main():
    """Run async MCP server"""
    logging.basicConfig(level=logging.INFO)
    server = AsyncMCPServer()

    # Use stdio transport for Claude Desktop
    async with stdio_server() as (read_stream, write_stream):
        await server.server.run(
            read_stream,
            write_stream,
            server.server.create_initialization_options()
        )

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

Quick Multi-Language MCP Patterns (Inlined for Standalone Use)

TypeScript vs Python Trade-offs:

Feature	TypeScript	Python
Type Safety	Compile-time + runtime (Zod)	Runtime only (Pydantic)
Performance	Faster startup, Node.js overhead	Slower startup, better for CPU tasks
Async Support	Native async/await, event loop	asyncio, great for I/O
Ecosystem	npm packages, frontend tools	Data science, ML libraries
Best For	Web APIs, real-time tools	Data processing, ML integration

Common Patterns Across Both:

1. Input Validation

   • TypeScript: Zod schemas
   • Python: Pydantic models

2. Error Handling

   • Both: Try/catch with specific error types
   • Return error content in MCP response

3. Resource Management

   • TypeScript: async/await with try/finally
   • Python: async context managers

4. Testing

   • TypeScript: Vitest/Jest with mock transport
   • Python: pytest with pytest-asyncio

Choosing Implementation Language:

// TypeScript - Best for:
// - File system operations
// - Web scraping/HTTP requests
// - JSON/API manipulation
// - Real-time data streams

// Python - Best for:
// - Data analysis (pandas, numpy)
// - Machine learning (scikit-learn, torch)
// - Database ETL operations
// - Scientific computing

[Full TypeScript, Python async, and OpenRouter patterns available in respective skills if deployed together]

Local MCP Integration Patterns (Your Repos)

.mcp.json Layout (Project-Scoped)

{
  "mcpServers": {
    "kuzu-memory": {
      "type": "stdio",
      "command": "kuzu-memory",
      "args": ["mcp"]
    },
    "mcp-vector-search": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "mcp-vector-search", "mcp"],
      "env": {
        "MCP_ENABLE_FILE_WATCHING": "true"
      }
    }
  }
}

CLI Conventions

• Provide a mcp subcommand for stdio servers.
• Use setup for end-to-end init + integration (mcp-vector-search, kuzu-memory).
• Use install / uninstall to target specific clients (mcp-ticketer).
• Provide doctor commands to validate dependencies.

Environment Conventions

• File watching: MCP_ENABLE_FILE_WATCHING=true
• Kuzu memory: KUZU_MEMORY_PROJECT_ROOT, KUZU_MEMORY_DB
• Ticketer adapters: MCP_TICKETER_ADAPTER, GITHUB_TOKEN, GITHUB_OWNER, GITHUB_REPO

Common Server Names

kuzu-memory, mcp-vector-search, mcp-ticketer, mcp-skillset, mcp-browser

Summary

• MCP enables AI-native server development with tools, resources, and prompts
• Tools are functions LLMs can call (read files, query APIs, run operations)
• Resources provide read-only data access (files, databases, documentation)
• Prompts are reusable templates with arguments for common tasks
• SDKs available for TypeScript and Python with full async support
• Claude Desktop integration via JSON config (STDIO transport)
• Debugging with MCP Inspector, server logs, and standalone testing
• Security critical: validate inputs, sanitize paths, use allowlists
• Best practices: Clear naming, comprehensive schemas, error handling, performance optimization