Skip to content

Latest commit

 

History

History
2236 lines (1772 loc) · 77.9 KB

File metadata and controls

2236 lines (1772 loc) · 77.9 KB

MCP開発のベストプラクティス

MCP開発のベストプラクティス

(上の画像をクリックすると、このレッスンの動画が視聴できます)

概要

このレッスンでは、MCPサーバーや機能を本番環境で開発、テスト、デプロイする際の高度なベストプラクティスに焦点を当てます。MCPエコシステムが複雑化し重要性が増すにつれ、確立されたパターンに従うことで信頼性、保守性、相互運用性を確保できます。このレッスンでは、実際のMCP実装から得られた実践的な知識を統合し、効率的で堅牢なサーバーを作成するためのリソース、プロンプト、ツールを提供します。

学習目標

このレッスンを終える頃には、以下ができるようになります:

  • MCPサーバーと機能設計における業界のベストプラクティスを適用する
  • MCPサーバーの包括的なテスト戦略を作成する
  • 複雑なMCPアプリケーション向けの効率的で再利用可能なワークフローパターンを設計する
  • MCPサーバーで適切なエラーハンドリング、ログ記録、可観測性を実装する
  • パフォーマンス、セキュリティ、保守性を最適化したMCP実装を行う

MCPの基本原則

具体的な実装のプラクティスに入る前に、効果的なMCP開発を導く基本原則を理解することが重要です:

  1. 標準化された通信: MCPはJSON-RPC 2.0を基盤としており、すべての実装でリクエスト、レスポンス、エラーハンドリングの一貫したフォーマットを提供します。

  2. ユーザー中心の設計: MCP実装では常にユーザーの同意、コントロール、透明性を優先してください。

  3. セキュリティ第一: 認証、認可、検証、レート制限を含む堅牢なセキュリティ対策を実施してください。

  4. モジュール型アーキテクチャ: 各ツールやリソースが明確で焦点を絞った目的を持つように、モジュール型のアプローチでMCPサーバーを設計してください。

  5. ステートフルな接続: MCPの機能を活用して複数のリクエスト間で状態を維持し、より一貫性のあるコンテキスト対応のインタラクションを実現してください。

MCP公式ベストプラクティス

以下のベストプラクティスは、公式のModel Context Protocolドキュメントから派生したものです:

セキュリティのベストプラクティス

  1. ユーザーの同意とコントロール: データにアクセスしたり操作を実行する前に、必ず明示的なユーザーの同意を求めてください。共有されるデータや許可されるアクションについて明確なコントロールを提供してください。

  2. データプライバシー: 明示的な同意なしにユーザーデータを公開せず、適切なアクセス制御で保護してください。不正なデータ送信を防ぐ措置を講じてください。

  3. ツールの安全性: どのツールを呼び出す場合でも、必ず明示的なユーザーの同意を求めてください。各ツールの機能をユーザーが理解できるようにし、堅牢なセキュリティ境界を確保してください。

  4. ツールの許可制御: セッション中にモデルが使用できるツールを設定し、明示的に許可されたツールのみがアクセス可能であることを確認してください。

  5. 認証: APIキー、OAuthトークン、その他の安全な認証方法を使用して、ツール、リソース、または機密操作へのアクセスを許可する前に適切な認証を要求してください。

  6. パラメータ検証: ツールの呼び出しに対して検証を強制し、不正な入力や悪意のある入力がツールの実装に到達するのを防いでください。

  7. レート制限: サーバーリソースの乱用を防ぎ、公平な使用を確保するためにレート制限を実施してください。

実装のベストプラクティス

  1. 機能交渉: 接続設定時に、サポートされている機能、プロトコルバージョン、利用可能なツールやリソースに関する情報を交換してください。

  2. ツール設計: 複数の懸念事項を扱うモノリシックなツールではなく、特定のタスクに特化したツールを作成してください。

  3. エラーハンドリング: 問題を診断し、失敗を優雅に処理し、実行可能なフィードバックを提供するために、標準化されたエラーメッセージとコードを実装してください。

  4. ログ記録: プロトコルインタラクションの監査、デバッグ、モニタリングのために構造化されたログを設定してください。

  5. 進捗追跡: 長時間実行される操作については、進捗状況を報告し、応答性の高いユーザーインターフェースを可能にしてください。

  6. リクエストのキャンセル: 必要なくなったり時間がかかりすぎたりするリクエストをクライアントがキャンセルできるようにしてください。

追加の参考資料

MCPのベストプラクティスに関する最新情報については、以下を参照してください:

実践的な実装例

ツール設計のベストプラクティス

1. 単一責任の原則

各MCPツールは明確で焦点を絞った目的を持つべきです。複数の懸念事項を扱おうとするモノリシックなツールを作成するのではなく、特定のタスクに特化したツールを開発してください。

// A focused tool that does one thing well
public class WeatherForecastTool : ITool
{
    private readonly IWeatherService _weatherService;
    
    public WeatherForecastTool(IWeatherService weatherService)
    {
        _weatherService = weatherService;
    }
    
    public string Name => "weatherForecast";
    public string Description => "Gets weather forecast for a specific location";
    
    public ToolDefinition GetDefinition()
    {
        return new ToolDefinition
        {
            Name = Name,
            Description = Description,
            Parameters = new Dictionary<string, ParameterDefinition>
            {
                ["location"] = new ParameterDefinition
                {
                    Type = ParameterType.String,
                    Description = "City or location name"
                },
                ["days"] = new ParameterDefinition
                {
                    Type = ParameterType.Integer,
                    Description = "Number of forecast days",
                    Default = 3
                }
            },
            Required = new[] { "location" }
        };
    }
    
    public async Task<ToolResponse> ExecuteAsync(IDictionary<string, object> parameters)
    {
        var location = parameters["location"].ToString();
        var days = parameters.ContainsKey("days") 
            ? Convert.ToInt32(parameters["days"]) 
            : 3;
            
        var forecast = await _weatherService.GetForecastAsync(location, days);
        
        return new ToolResponse
        {
            Content = new List<ContentItem>
            {
                new TextContent(JsonSerializer.Serialize(forecast))
            }
        };
    }
}

2. 一貫したエラーハンドリング

情報豊富なエラーメッセージと適切な回復メカニズムを備えた堅牢なエラーハンドリングを実装してください。

# Python example with comprehensive error handling
class DataQueryTool:
    def get_name(self):
        return "dataQuery"
        
    def get_description(self):
        return "Queries data from specified database tables"
    
    async def execute(self, parameters):
        try:
            # Parameter validation
            if "query" not in parameters:
                raise ToolParameterError("Missing required parameter: query")
                
            query = parameters["query"]
            
            # Security validation
            if self._contains_unsafe_sql(query):
                raise ToolSecurityError("Query contains potentially unsafe SQL")
            
            try:
                # Database operation with timeout
                async with timeout(10):  # 10 second timeout
                    result = await self._database.execute_query(query)
                    
                return ToolResponse(
                    content=[TextContent(json.dumps(result))]
                )
            except asyncio.TimeoutError:
                raise ToolExecutionError("Database query timed out after 10 seconds")
            except DatabaseConnectionError as e:
                # Connection errors might be transient
                self._log_error("Database connection error", e)
                raise ToolExecutionError(f"Database connection error: {str(e)}")
            except DatabaseQueryError as e:
                # Query errors are likely client errors
                self._log_error("Database query error", e)
                raise ToolExecutionError(f"Invalid query: {str(e)}")
                
        except ToolError:
            # Let tool-specific errors pass through
            raise
        except Exception as e:
            # Catch-all for unexpected errors
            self._log_error("Unexpected error in DataQueryTool", e)
            raise ToolExecutionError(f"An unexpected error occurred: {str(e)}")
    
    def _contains_unsafe_sql(self, query):
        # Implementation of SQL injection detection
        pass
        
    def _log_error(self, message, error):
        # Implementation of error logging
        pass

3. パラメータ検証

不正な入力や悪意のある入力を防ぐために、常にパラメータを徹底的に検証してください。

// JavaScript/TypeScript example with detailed parameter validation
class FileOperationTool {
  getName() {
    return "fileOperation";
  }
  
  getDescription() {
    return "Performs file operations like read, write, and delete";
  }
  
  getDefinition() {
    return {
      name: this.getName(),
      description: this.getDescription(),
      parameters: {
        operation: {
          type: "string",
          description: "Operation to perform",
          enum: ["read", "write", "delete"]
        },
        path: {
          type: "string",
          description: "File path (must be within allowed directories)"
        },
        content: {
          type: "string",
          description: "Content to write (only for write operation)",
          optional: true
        }
      },
      required: ["operation", "path"]
    };
  }
  
  async execute(parameters) {
    // 1. Validate parameter presence
    if (!parameters.operation) {
      throw new ToolError("Missing required parameter: operation");
    }
    
    if (!parameters.path) {
      throw new ToolError("Missing required parameter: path");
    }
    
    // 2. Validate parameter types
    if (typeof parameters.operation !== "string") {
      throw new ToolError("Parameter 'operation' must be a string");
    }
    
    if (typeof parameters.path !== "string") {
      throw new ToolError("Parameter 'path' must be a string");
    }
    
    // 3. Validate parameter values
    const validOperations = ["read", "write", "delete"];
    if (!validOperations.includes(parameters.operation)) {
      throw new ToolError(`Invalid operation. Must be one of: ${validOperations.join(", ")}`);
    }
    
    // 4. Validate content presence for write operation
    if (parameters.operation === "write" && !parameters.content) {
      throw new ToolError("Content parameter is required for write operation");
    }
    
    // 5. Path safety validation
    if (!this.isPathWithinAllowedDirectories(parameters.path)) {
      throw new ToolError("Access denied: path is outside of allowed directories");
    }
    
    // Implementation based on validated parameters
    // ...
  }
  
  isPathWithinAllowedDirectories(path) {
    // Implementation of path safety check
    // ...
  }
}

セキュリティ実装例

1. 認証と認可

// Java example with authentication and authorization
public class SecureDataAccessTool implements Tool {
    private final AuthenticationService authService;
    private final AuthorizationService authzService;
    private final DataService dataService;
    
    // Dependency injection
    public SecureDataAccessTool(
            AuthenticationService authService,
            AuthorizationService authzService,
            DataService dataService) {
        this.authService = authService;
        this.authzService = authzService;
        this.dataService = dataService;
    }
    
    @Override
    public String getName() {
        return "secureDataAccess";
    }
    
    @Override
    public ToolResponse execute(ToolRequest request) {
        // 1. Extract authentication context
        String authToken = request.getContext().getAuthToken();
        
        // 2. Authenticate user
        UserIdentity user;
        try {
            user = authService.validateToken(authToken);
        } catch (AuthenticationException e) {
            return ToolResponse.error("Authentication failed: " + e.getMessage());
        }
        
        // 3. Check authorization for the specific operation
        String dataId = request.getParameters().get("dataId").getAsString();
        String operation = request.getParameters().get("operation").getAsString();
        
        boolean isAuthorized = authzService.isAuthorized(user, "data:" + dataId, operation);
        if (!isAuthorized) {
            return ToolResponse.error("Access denied: Insufficient permissions for this operation");
        }
        
        // 4. Proceed with authorized operation
        try {
            switch (operation) {
                case "read":
                    Object data = dataService.getData(dataId, user.getId());
                    return ToolResponse.success(data);
                case "update":
                    JsonNode newData = request.getParameters().get("newData");
                    dataService.updateData(dataId, newData, user.getId());
                    return ToolResponse.success("Data updated successfully");
                default:
                    return ToolResponse.error("Unsupported operation: " + operation);
            }
        } catch (Exception e) {
            return ToolResponse.error("Operation failed: " + e.getMessage());
        }
    }
}

2. レート制限

// C# rate limiting implementation
public class RateLimitingMiddleware
{
    private readonly RequestDelegate _next;
    private readonly IMemoryCache _cache;
    private readonly ILogger<RateLimitingMiddleware> _logger;
    
    // Configuration options
    private readonly int _maxRequestsPerMinute;
    
    public RateLimitingMiddleware(
        RequestDelegate next,
        IMemoryCache cache,
        ILogger<RateLimitingMiddleware> logger,
        IConfiguration config)
    {
        _next = next;
        _cache = cache;
        _logger = logger;
        _maxRequestsPerMinute = config.GetValue<int>("RateLimit:MaxRequestsPerMinute", 60);
    }
    
    public async Task InvokeAsync(HttpContext context)
    {
        // 1. Get client identifier (API key or user ID)
        string clientId = GetClientIdentifier(context);
        
        // 2. Get rate limiting key for this minute
        string cacheKey = $"rate_limit:{clientId}:{DateTime.UtcNow:yyyyMMddHHmm}";
        
        // 3. Check current request count
        if (!_cache.TryGetValue(cacheKey, out int requestCount))
        {
            requestCount = 0;
        }
        
        // 4. Enforce rate limit
        if (requestCount >= _maxRequestsPerMinute)
        {
            _logger.LogWarning("Rate limit exceeded for client {ClientId}", clientId);
            
            context.Response.StatusCode = StatusCodes.Status429TooManyRequests;
            context.Response.Headers.Add("Retry-After", "60");
            
            await context.Response.WriteAsJsonAsync(new
            {
                error = "Rate limit exceeded",
                message = "Too many requests. Please try again later.",
                retryAfterSeconds = 60
            });
            
            return;
        }
        
        // 5. Increment request count
        _cache.Set(cacheKey, requestCount + 1, TimeSpan.FromMinutes(2));
        
        // 6. Add rate limit headers
        context.Response.Headers.Add("X-RateLimit-Limit", _maxRequestsPerMinute.ToString());
        context.Response.Headers.Add("X-RateLimit-Remaining", (_maxRequestsPerMinute - requestCount - 1).ToString());
        
        // 7. Continue with the request
        await _next(context);
    }
    
    private string GetClientIdentifier(HttpContext context)
    {
        // Implementation to extract API key or user ID
        // ...
    }
}

テストのベストプラクティス

1. MCPツールの単体テスト

外部依存関係をモックしてツールを単独でテストしてください:

// TypeScript example of a tool unit test
describe('WeatherForecastTool', () => {
  let tool: WeatherForecastTool;
  let mockWeatherService: jest.Mocked<IWeatherService>;
  
  beforeEach(() => {
    // Create a mock weather service
    mockWeatherService = {
      getForecasts: jest.fn()
    } as any;
    
    // Create the tool with the mock dependency
    tool = new WeatherForecastTool(mockWeatherService);
  });
  
  it('should return weather forecast for a location', async () => {
    // Arrange
    const mockForecast = {
      location: 'Seattle',
      forecasts: [
        { date: '2025-07-16', temperature: 72, conditions: 'Sunny' },
        { date: '2025-07-17', temperature: 68, conditions: 'Partly Cloudy' },
        { date: '2025-07-18', temperature: 65, conditions: 'Rain' }
      ]
    };
    
    mockWeatherService.getForecasts.mockResolvedValue(mockForecast);
    
    // Act
    const response = await tool.execute({
      location: 'Seattle',
      days: 3
    });
    
    // Assert
    expect(mockWeatherService.getForecasts).toHaveBeenCalledWith('Seattle', 3);
    expect(response.content[0].text).toContain('Seattle');
    expect(response.content[0].text).toContain('Sunny');
  });
  
  it('should handle errors from the weather service', async () => {
    // Arrange
    mockWeatherService.getForecasts.mockRejectedValue(new Error('Service unavailable'));
    
    // Act & Assert
    await expect(tool.execute({
      location: 'Seattle',
      days: 3
    })).rejects.toThrow('Weather service error: Service unavailable');
  });
});

2. 統合テスト

クライアントリクエストからサーバーレスポンスまでの完全なフローをテストしてください:

# Python integration test example
@pytest.mark.asyncio
async def test_mcp_server_integration():
    # Start a test server
    server = McpServer()
    server.register_tool(WeatherForecastTool(MockWeatherService()))
    await server.start(port=5000)
    
    try:
        # Create a client
        client = McpClient("http://localhost:5000")
        
        # Test tool discovery
        tools = await client.discover_tools()
        assert "weatherForecast" in [t.name for t in tools]
        
        # Test tool execution
        response = await client.execute_tool("weatherForecast", {
            "location": "Seattle",
            "days": 3
        })
        
        # Verify response
        assert response.status_code == 200
        assert "Seattle" in response.content[0].text
        assert len(json.loads(response.content[0].text)["forecasts"]) == 3
        
    finally:
        # Clean up
        await server.stop()

パフォーマンス最適化

1. キャッシュ戦略

遅延とリソース使用量を削減するために適切なキャッシュを実装してください:

// C# example with caching
public class CachedWeatherTool : ITool
{
    private readonly IWeatherService _weatherService;
    private readonly IDistributedCache _cache;
    private readonly ILogger<CachedWeatherTool> _logger;
    
    public CachedWeatherTool(
        IWeatherService weatherService,
        IDistributedCache cache,
        ILogger<CachedWeatherTool> logger)
    {
        _weatherService = weatherService;
        _cache = cache;
        _logger = logger;
    }
    
    public string Name => "weatherForecast";
    
    public async Task<ToolResponse> ExecuteAsync(IDictionary<string, object> parameters)
    {
        var location = parameters["location"].ToString();
        var days = Convert.ToInt32(parameters.GetValueOrDefault("days", 3));
        
        // Create cache key
        string cacheKey = $"weather:{location}:{days}";
        
        // Try to get from cache
        string cachedForecast = await _cache.GetStringAsync(cacheKey);
        if (!string.IsNullOrEmpty(cachedForecast))
        {
            _logger.LogInformation("Cache hit for weather forecast: {Location}", location);
            return new ToolResponse
            {
                Content = new List<ContentItem>
                {
                    new TextContent(cachedForecast)
                }
            };
        }
        
        // Cache miss - get from service
        _logger.LogInformation("Cache miss for weather forecast: {Location}", location);
        var forecast = await _weatherService.GetForecastAsync(location, days);
        string forecastJson = JsonSerializer.Serialize(forecast);
        
        // Store in cache (weather forecasts valid for 1 hour)
        await _cache.SetStringAsync(
            cacheKey,
            forecastJson,
            new DistributedCacheEntryOptions
            {
                AbsoluteExpirationRelativeToNow = TimeSpan.FromHours(1)
            });
        
        return new ToolResponse
        {
            Content = new List<ContentItem>
            {
                new TextContent(forecastJson)
            }
        };
    }
}

2. 依存性注入とテスト可能性

ツールがコンストラクタ注入を通じて依存関係を受け取るように設計し、テスト可能で構成可能にしてください:

// Java example with dependency injection
public class CurrencyConversionTool implements Tool {
    private final ExchangeRateService exchangeService;
    private final CacheService cacheService;
    private final Logger logger;
    
    // Dependencies injected through constructor
    public CurrencyConversionTool(
            ExchangeRateService exchangeService,
            CacheService cacheService,
            Logger logger) {
        this.exchangeService = exchangeService;
        this.cacheService = cacheService;
        this.logger = logger;
    }
    
    // Tool implementation
    // ...
}

3. 合成可能なツール

より複雑なワークフローを作成するためにツールを組み合わせて設計してください:

# Python example showing composable tools
class DataFetchTool(Tool):
    def get_name(self):
        return "dataFetch"
    
    # Implementation...

class DataAnalysisTool(Tool):
    def get_name(self):
        return "dataAnalysis"
    
    # This tool can use results from the dataFetch tool
    async def execute_async(self, request):
        # Implementation...
        pass

class DataVisualizationTool(Tool):
    def get_name(self):
        return "dataVisualize"
    
    # This tool can use results from the dataAnalysis tool
    async def execute_async(self, request):
        # Implementation...
        pass

# These tools can be used independently or as part of a workflow

スキーマ設計のベストプラクティス

スキーマはモデルとツール間の契約です。よく設計されたスキーマはツールの使いやすさを向上させます。

1. 明確なパラメータ説明

各パラメータに説明情報を必ず含めてください:

public object GetSchema()
{
    return new {
        type = "object",
        properties = new {
            query = new { 
                type = "string", 
                description = "Search query text. Use precise keywords for better results." 
            },
            filters = new {
                type = "object",
                description = "Optional filters to narrow down search results",
                properties = new {
                    dateRange = new { 
                        type = "string", 
                        description = "Date range in format YYYY-MM-DD:YYYY-MM-DD" 
                    },
                    category = new { 
                        type = "string", 
                        description = "Category name to filter by" 
                    }
                }
            },
            limit = new { 
                type = "integer", 
                description = "Maximum number of results to return (1-50)",
                default = 10
            }
        },
        required = new[] { "query" }
    };
}

2. 検証制約

無効な入力を防ぐために検証制約を含めてください:

Map<String, Object> getSchema() {
    Map<String, Object> schema = new HashMap<>();
    schema.put("type", "object");
    
    Map<String, Object> properties = new HashMap<>();
    
    // Email property with format validation
    Map<String, Object> email = new HashMap<>();
    email.put("type", "string");
    email.put("format", "email");
    email.put("description", "User email address");
    
    // Age property with numeric constraints
    Map<String, Object> age = new HashMap<>();
    age.put("type", "integer");
    age.put("minimum", 13);
    age.put("maximum", 120);
    age.put("description", "User age in years");
    
    // Enumerated property
    Map<String, Object> subscription = new HashMap<>();
    subscription.put("type", "string");
    subscription.put("enum", Arrays.asList("free", "basic", "premium"));
    subscription.put("default", "free");
    subscription.put("description", "Subscription tier");
    
    properties.put("email", email);
    properties.put("age", age);
    properties.put("subscription", subscription);
    
    schema.put("properties", properties);
    schema.put("required", Arrays.asList("email"));
    
    return schema;
}

3. 一貫した返却構造

モデルが結果を解釈しやすくするために、レスポンス構造の一貫性を維持してください:

async def execute_async(self, request):
    try:
        # Process request
        results = await self._search_database(request.parameters["query"])
        
        # Always return a consistent structure
        return ToolResponse(
            result={
                "matches": [self._format_item(item) for item in results],
                "totalCount": len(results),
                "queryTime": calculation_time_ms,
                "status": "success"
            }
        )
    except Exception as e:
        return ToolResponse(
            result={
                "matches": [],
                "totalCount": 0,
                "queryTime": 0,
                "status": "error",
                "error": str(e)
            }
        )
    
def _format_item(self, item):
    """Ensures each item has a consistent structure"""
    return {
        "id": item.id,
        "title": item.title,
        "summary": item.summary[:100] + "..." if len(item.summary) > 100 else item.summary,
        "url": item.url,
        "relevance": item.score
    }

エラーハンドリング

MCPツールの信頼性を維持するためには、堅牢なエラーハンドリングが不可欠です。

1. 優雅なエラーハンドリング

適切なレベルでエラーを処理し、情報豊富なメッセージを提供してください:

public async Task<ToolResponse> ExecuteAsync(ToolRequest request)
{
    try
    {
        string fileId = request.Parameters.GetProperty("fileId").GetString();
        
        try
        {
            var fileData = await _fileService.GetFileAsync(fileId);
            return new ToolResponse { 
                Result = JsonSerializer.SerializeToElement(fileData) 
            };
        }
        catch (FileNotFoundException)
        {
            throw new ToolExecutionException($"File not found: {fileId}");
        }
        catch (UnauthorizedAccessException)
        {
            throw new ToolExecutionException("You don't have permission to access this file");
        }
        catch (Exception ex) when (ex is IOException || ex is TimeoutException)
        {
            _logger.LogError(ex, "Error accessing file {FileId}", fileId);
            throw new ToolExecutionException("Error accessing file: The service is temporarily unavailable");
        }
    }
    catch (JsonException)
    {
        throw new ToolExecutionException("Invalid file ID format");
    }
    catch (Exception ex)
    {
        _logger.LogError(ex, "Unexpected error in FileAccessTool");
        throw new ToolExecutionException("An unexpected error occurred");
    }
}

2. 構造化されたエラーレスポンス

可能であれば構造化されたエラー情報を返してください:

@Override
public ToolResponse execute(ToolRequest request) {
    try {
        // Implementation
    } catch (Exception ex) {
        Map<String, Object> errorResult = new HashMap<>();
        
        errorResult.put("success", false);
        
        if (ex instanceof ValidationException) {
            ValidationException validationEx = (ValidationException) ex;
            
            errorResult.put("errorType", "validation");
            errorResult.put("errorMessage", validationEx.getMessage());
            errorResult.put("validationErrors", validationEx.getErrors());
            
            return new ToolResponse.Builder()
                .setResult(errorResult)
                .build();
        }
        
        // Re-throw other exceptions as ToolExecutionException
        throw new ToolExecutionException("Tool execution failed: " + ex.getMessage(), ex);
    }
}

3. リトライロジック

一時的な失敗に対して適切なリトライロジックを実装してください:

async def execute_async(self, request):
    max_retries = 3
    retry_count = 0
    base_delay = 1  # seconds
    
    while retry_count < max_retries:
        try:
            # Call external API
            return await self._call_api(request.parameters)
        except TransientError as e:
            retry_count += 1
            if retry_count >= max_retries:
                raise ToolExecutionException(f"Operation failed after {max_retries} attempts: {str(e)}")
                
            # Exponential backoff
            delay = base_delay * (2 ** (retry_count - 1))
            logging.warning(f"Transient error, retrying in {delay}s: {str(e)}")
            await asyncio.sleep(delay)
        except Exception as e:
            # Non-transient error, don't retry
            raise ToolExecutionException(f"Operation failed: {str(e)}")

パフォーマンス最適化

1. キャッシュ

高コストな操作に対してキャッシュを実装してください:

public class CachedDataTool : IMcpTool
{
    private readonly IDatabase _database;
    private readonly IMemoryCache _cache;
    
    public CachedDataTool(IDatabase database, IMemoryCache cache)
    {
        _database = database;
        _cache = cache;
    }
    
    public async Task<ToolResponse> ExecuteAsync(ToolRequest request)
    {
        var query = request.Parameters.GetProperty("query").GetString();
        
        // Create cache key based on parameters
        var cacheKey = $"data_query_{ComputeHash(query)}";
        
        // Try to get from cache first
        if (_cache.TryGetValue(cacheKey, out var cachedResult))
        {
            return new ToolResponse { Result = cachedResult };
        }
        
        // Cache miss - perform actual query
        var result = await _database.QueryAsync(query);
        
        // Store in cache with expiration
        var cacheOptions = new MemoryCacheEntryOptions()
            .SetAbsoluteExpiration(TimeSpan.FromMinutes(15));
            
        _cache.Set(cacheKey, JsonSerializer.SerializeToElement(result), cacheOptions);
        
        return new ToolResponse { Result = JsonSerializer.SerializeToElement(result) };
    }
    
    private string ComputeHash(string input)
    {
        // Implementation to generate stable hash for cache key
    }
}

2. 非同期処理

I/Oバウンド操作に対して非同期プログラミングパターンを使用してください:

public class AsyncDocumentProcessingTool implements Tool {
    private final DocumentService documentService;
    private final ExecutorService executorService;
    
    @Override
    public ToolResponse execute(ToolRequest request) {
        String documentId = request.getParameters().get("documentId").asText();
        
        // For long-running operations, return a processing ID immediately
        String processId = UUID.randomUUID().toString();
        
        // Start async processing
        CompletableFuture.runAsync(() -> {
            try {
                // Perform long-running operation
                documentService.processDocument(documentId);
                
                // Update status (would typically be stored in a database)
                processStatusRepository.updateStatus(processId, "completed");
            } catch (Exception ex) {
                processStatusRepository.updateStatus(processId, "failed", ex.getMessage());
            }
        }, executorService);
        
        // Return immediate response with process ID
        Map<String, Object> result = new HashMap<>();
        result.put("processId", processId);
        result.put("status", "processing");
        result.put("estimatedCompletionTime", ZonedDateTime.now().plusMinutes(5));
        
        return new ToolResponse.Builder().setResult(result).build();
    }
    
    // Companion status check tool
    public class ProcessStatusTool implements Tool {
        @Override
        public ToolResponse execute(ToolRequest request) {
            String processId = request.getParameters().get("processId").asText();
            ProcessStatus status = processStatusRepository.getStatus(processId);
            
            return new ToolResponse.Builder().setResult(status).build();
        }
    }
}

3. リソーススロットリング

過負荷を防ぐためにリソーススロットリングを実装してください:

class ThrottledApiTool(Tool):
    def __init__(self):
        self.rate_limiter = TokenBucketRateLimiter(
            tokens_per_second=5,  # Allow 5 requests per second
            bucket_size=10        # Allow bursts up to 10 requests
        )
    
    async def execute_async(self, request):
        # Check if we can proceed or need to wait
        delay = self.rate_limiter.get_delay_time()
        
        if delay > 0:
            if delay > 2.0:  # If wait is too long
                raise ToolExecutionException(
                    f"Rate limit exceeded. Please try again in {delay:.1f} seconds."
                )
            else:
                # Wait for the appropriate delay time
                await asyncio.sleep(delay)
        
        # Consume a token and proceed with the request
        self.rate_limiter.consume()
        
        # Call API
        result = await self._call_api(request.parameters)
        return ToolResponse(result=result)

class TokenBucketRateLimiter:
    def __init__(self, tokens_per_second, bucket_size):
        self.tokens_per_second = tokens_per_second
        self.bucket_size = bucket_size
        self.tokens = bucket_size
        self.last_refill = time.time()
        self.lock = asyncio.Lock()
    
    async def get_delay_time(self):
        async with self.lock:
            self._refill()
            if self.tokens >= 1:
                return 0
            
            # Calculate time until next token available
            return (1 - self.tokens) / self.tokens_per_second
    
    async def consume(self):
        async with self.lock:
            self._refill()
            self.tokens -= 1
    
    def _refill(self):
        now = time.time()
        elapsed = now - self.last_refill
        
        # Add new tokens based on elapsed time
        new_tokens = elapsed * self.tokens_per_second
        self.tokens = min(self.bucket_size, self.tokens + new_tokens)
        self.last_refill = now

セキュリティのベストプラクティス

1. 入力検証

常に入力パラメータを徹底的に検証してください:

public async Task<ToolResponse> ExecuteAsync(ToolRequest request)
{
    // Validate parameters exist
    if (!request.Parameters.TryGetProperty("query", out var queryProp))
    {
        throw new ToolExecutionException("Missing required parameter: query");
    }
    
    // Validate correct type
    if (queryProp.ValueKind != JsonValueKind.String)
    {
        throw new ToolExecutionException("Query parameter must be a string");
    }
    
    var query = queryProp.GetString();
    
    // Validate string content
    if (string.IsNullOrWhiteSpace(query))
    {
        throw new ToolExecutionException("Query parameter cannot be empty");
    }
    
    if (query.Length > 500)
    {
        throw new ToolExecutionException("Query parameter exceeds maximum length of 500 characters");
    }
    
    // Check for SQL injection attacks if applicable
    if (ContainsSqlInjection(query))
    {
        throw new ToolExecutionException("Invalid query: contains potentially unsafe SQL");
    }
    
    // Proceed with execution
    // ...
}

2. 認可チェック

適切な認可チェックを実装してください:

@Override
public ToolResponse execute(ToolRequest request) {
    // Get user context from request
    UserContext user = request.getContext().getUserContext();
    
    // Check if user has required permissions
    if (!authorizationService.hasPermission(user, "documents:read")) {
        throw new ToolExecutionException("User does not have permission to access documents");
    }
    
    // For specific resources, check access to that resource
    String documentId = request.getParameters().get("documentId").asText();
    if (!documentService.canUserAccess(user.getId(), documentId)) {
        throw new ToolExecutionException("Access denied to the requested document");
    }
    
    // Proceed with tool execution
    // ...
}

3. 機密データの取り扱い

機密データを慎重に取り扱ってください:

class SecureDataTool(Tool):
    def get_schema(self):
        return {
            "type": "object",
            "properties": {
                "userId": {"type": "string"},
                "includeSensitiveData": {"type": "boolean", "default": False}
            },
            "required": ["userId"]
        }
    
    async def execute_async(self, request):
        user_id = request.parameters["userId"]
        include_sensitive = request.parameters.get("includeSensitiveData", False)
        
        # Get user data
        user_data = await self.user_service.get_user_data(user_id)
        
        # Filter sensitive fields unless explicitly requested AND authorized
        if not include_sensitive or not self._is_authorized_for_sensitive_data(request):
            user_data = self._redact_sensitive_fields(user_data)
        
        return ToolResponse(result=user_data)
    
    def _is_authorized_for_sensitive_data(self, request):
        # Check authorization level in request context
        auth_level = request.context.get("authorizationLevel")
        return auth_level == "admin"
    
    def _redact_sensitive_fields(self, user_data):
        # Create a copy to avoid modifying the original
        redacted = user_data.copy()
        
        # Redact specific sensitive fields
        sensitive_fields = ["ssn", "creditCardNumber", "password"]
        for field in sensitive_fields:
            if field in redacted:
                redacted[field] = "REDACTED"
        
        # Redact nested sensitive data
        if "financialInfo" in redacted:
            redacted["financialInfo"] = {"available": True, "accessRestricted": True}
        
        return redacted

MCPツールのテストのベストプラクティス

包括的なテストは、MCPツールが正しく機能し、エッジケースを処理し、システム全体と適切に統合されることを保証します。

単体テスト

1. 各ツールを単独でテスト

各ツールの機能に焦点を当てたテストを作成してください:

[Fact]
public async Task WeatherTool_ValidLocation_ReturnsCorrectForecast()
{
    // Arrange
    var mockWeatherService = new Mock<IWeatherService>();
    mockWeatherService
        .Setup(s => s.GetForecastAsync("Seattle", 3))
        .ReturnsAsync(new WeatherForecast(/* test data */));
    
    var tool = new WeatherForecastTool(mockWeatherService.Object);
    
    var request = new ToolRequest(
        toolName: "weatherForecast",
        parameters: JsonSerializer.SerializeToElement(new { 
            location = "Seattle", 
            days = 3 
        })
    );
    
    // Act
    var response = await tool.ExecuteAsync(request);
    
    // Assert
    Assert.NotNull(response);
    var result = JsonSerializer.Deserialize<WeatherForecast>(response.Result);
    Assert.Equal("Seattle", result.Location);
    Assert.Equal(3, result.DailyForecasts.Count);
}

[Fact]
public async Task WeatherTool_InvalidLocation_ThrowsToolExecutionException()
{
    // Arrange
    var mockWeatherService = new Mock<IWeatherService>();
    mockWeatherService
        .Setup(s => s.GetForecastAsync("InvalidLocation", It.IsAny<int>()))
        .ThrowsAsync(new LocationNotFoundException("Location not found"));
    
    var tool = new WeatherForecastTool(mockWeatherService.Object);
    
    var request = new ToolRequest(
        toolName: "weatherForecast",
        parameters: JsonSerializer.SerializeToElement(new { 
            location = "InvalidLocation", 
            days = 3 
        })
    );
    
    // Act & Assert
    var exception = await Assert.ThrowsAsync<ToolExecutionException>(
        () => tool.ExecuteAsync(request)
    );
    
    Assert.Contains("Location not found", exception.Message);
}

2. スキーマ検証テスト

スキーマが有効であり、適切に制約を強制することをテストしてください:

@Test
public void testSchemaValidation() {
    // Create tool instance
    SearchTool searchTool = new SearchTool();
    
    // Get schema
    Object schema = searchTool.getSchema();
    
    // Convert schema to JSON for validation
    String schemaJson = objectMapper.writeValueAsString(schema);
    
    // Validate schema is valid JSONSchema
    JsonSchemaFactory factory = JsonSchemaFactory.byDefault();
    JsonSchema jsonSchema = factory.getJsonSchema(schemaJson);
    
    // Test valid parameters
    JsonNode validParams = objectMapper.createObjectNode()
        .put("query", "test query")
        .put("limit", 5);
        
    ProcessingReport validReport = jsonSchema.validate(validParams);
    assertTrue(validReport.isSuccess());
    
    // Test missing required parameter
    JsonNode missingRequired = objectMapper.createObjectNode()
        .put("limit", 5);
        
    ProcessingReport missingReport = jsonSchema.validate(missingRequired);
    assertFalse(missingReport.isSuccess());
    
    // Test invalid parameter type
    JsonNode invalidType = objectMapper.createObjectNode()
        .put("query", "test")
        .put("limit", "not-a-number");
        
    ProcessingReport invalidReport = jsonSchema.validate(invalidType);
    assertFalse(invalidReport.isSuccess());
}

3. エラーハンドリングテスト

エラー条件に対する特定のテストを作成してください:

@pytest.mark.asyncio
async def test_api_tool_handles_timeout():
    # Arrange
    tool = ApiTool(timeout=0.1)  # Very short timeout
    
    # Mock a request that will time out
    with aioresponses() as mocked:
        mocked.get(
            "https://api.example.com/data",
            callback=lambda *args, **kwargs: asyncio.sleep(0.5)  # Longer than timeout
        )
        
        request = ToolRequest(
            tool_name="apiTool",
            parameters={"url": "https://api.example.com/data"}
        )
        
        # Act & Assert
        with pytest.raises(ToolExecutionException) as exc_info:
            await tool.execute_async(request)
        
        # Verify exception message
        assert "timed out" in str(exc_info.value).lower()

@pytest.mark.asyncio
async def test_api_tool_handles_rate_limiting():
    # Arrange
    tool = ApiTool()
    
    # Mock a rate-limited response
    with aioresponses() as mocked:
        mocked.get(
            "https://api.example.com/data",
            status=429,
            headers={"Retry-After": "2"},
            body=json.dumps({"error": "Rate limit exceeded"})
        )
        
        request = ToolRequest(
            tool_name="apiTool",
            parameters={"url": "https://api.example.com/data"}
        )
        
        # Act & Assert
        with pytest.raises(ToolExecutionException) as exc_info:
            await tool.execute_async(request)
        
        # Verify exception contains rate limit information
        error_msg = str(exc_info.value).lower()
        assert "rate limit" in error_msg
        assert "try again" in error_msg

統合テスト

1. ツールチェーンテスト

期待される組み合わせで動作するツールをテストしてください:

[Fact]
public async Task DataProcessingWorkflow_CompletesSuccessfully()
{
    // Arrange
    var dataFetchTool = new DataFetchTool(mockDataService.Object);
    var analysisTools = new DataAnalysisTool(mockAnalysisService.Object);
    var visualizationTool = new DataVisualizationTool(mockVisualizationService.Object);
    
    var toolRegistry = new ToolRegistry();
    toolRegistry.RegisterTool(dataFetchTool);
    toolRegistry.RegisterTool(analysisTools);
    toolRegistry.RegisterTool(visualizationTool);
    
    var workflowExecutor = new WorkflowExecutor(toolRegistry);
    
    // Act
    var result = await workflowExecutor.ExecuteWorkflowAsync(new[] {
        new ToolCall("dataFetch", new { source = "sales2023" }),
        new ToolCall("dataAnalysis", ctx => new { 
            data = ctx.GetResult("dataFetch"),
            analysis = "trend" 
        }),
        new ToolCall("dataVisualize", ctx => new {
            analysisResult = ctx.GetResult("dataAnalysis"),
            type = "line-chart"
        })
    });
    
    // Assert
    Assert.NotNull(result);
    Assert.True(result.Success);
    Assert.NotNull(result.GetResult("dataVisualize"));
    Assert.Contains("chartUrl", result.GetResult("dataVisualize").ToString());
}

2. MCPサーバーテスト

ツールの登録と実行を含む完全なMCPサーバーをテストしてください:

@SpringBootTest
@AutoConfigureMockMvc
public class McpServerIntegrationTest {
    
    @Autowired
    private MockMvc mockMvc;
    
    @Autowired
    private ObjectMapper objectMapper;
    
    @Test
    public void testToolDiscovery() throws Exception {
        // Test the discovery endpoint
        mockMvc.perform(get("/mcp/tools"))
            .andExpect(status().isOk())
            .andExpect(jsonPath("$.tools").isArray())
            .andExpect(jsonPath("$.tools[*].name").value(hasItems(
                "weatherForecast", "calculator", "documentSearch"
            )));
    }
    
    @Test
    public void testToolExecution() throws Exception {
        // Create tool request
        Map<String, Object> request = new HashMap<>();
        request.put("toolName", "calculator");
        
        Map<String, Object> parameters = new HashMap<>();
        parameters.put("operation", "add");
        parameters.put("a", 5);
        parameters.put("b", 7);
        request.put("parameters", parameters);
        
        // Send request and verify response
        mockMvc.perform(post("/mcp/execute")
            .contentType(MediaType.APPLICATION_JSON)
            .content(objectMapper.writeValueAsString(request)))
            .andExpect(status().isOk())
            .andExpect(jsonPath("$.result.value").value(12));
    }
    
    @Test
    public void testToolValidation() throws Exception {
        // Create invalid tool request
        Map<String, Object> request = new HashMap<>();
        request.put("toolName", "calculator");
        
        Map<String, Object> parameters = new HashMap<>();
        parameters.put("operation", "divide");
        parameters.put("a", 10);
        // Missing parameter "b"
        request.put("parameters", parameters);
        
        // Send request and verify error response
        mockMvc.perform(post("/mcp/execute")
            .contentType(MediaType.APPLICATION_JSON)
            .content(objectMapper.writeValueAsString(request)))
            .andExpect(status().isBadRequest())
            .andExpect(jsonPath("$.error").exists());
    }
}

3. エンドツーエンドテスト

モデルプロンプトからツール実行までの完全なワークフローをテストしてください:

@pytest.mark.asyncio
async def test_model_interaction_with_tool():
    # Arrange - Set up MCP client and mock model
    mcp_client = McpClient(server_url="http://localhost:5000")
    
    # Mock model responses
    mock_model = MockLanguageModel([
        MockResponse(
            "What's the weather in Seattle?",
            tool_calls=[{
                "tool_name": "weatherForecast",
                "parameters": {"location": "Seattle", "days": 3}
            }]
        ),
        MockResponse(
            "Here's the weather forecast for Seattle:\n- Today: 65°F, Partly Cloudy\n- Tomorrow: 68°F, Sunny\n- Day after: 62°F, Rain",
            tool_calls=[]
        )
    ])
    
    # Mock weather tool response
    with aioresponses() as mocked:
        mocked.post(
            "http://localhost:5000/mcp/execute",
            payload={
                "result": {
                    "location": "Seattle",
                    "forecast": [
                        {"date": "2023-06-01", "temperature": 65, "conditions": "Partly Cloudy"},
                        {"date": "2023-06-02", "temperature": 68, "conditions": "Sunny"},
                        {"date": "2023-06-03", "temperature": 62, "conditions": "Rain"}
                    ]
                }
            }
        )
        
        # Act
        response = await mcp_client.send_prompt(
            "What's the weather in Seattle?",
            model=mock_model,
            allowed_tools=["weatherForecast"]
        )
        
        # Assert
        assert "Seattle" in response.generated_text
        assert "65" in response.generated_text
        assert "Sunny" in response.generated_text
        assert "Rain" in response.generated_text
        assert len(response.tool_calls) == 1
        assert response.tool_calls[0].tool_name == "weatherForecast"

パフォーマンステスト

1. 負荷テスト

MCPサーバーが処理できる同時リクエスト数をテストしてください:

[Fact]
public async Task McpServer_HandlesHighConcurrency()
{
    // Arrange
    var server = new McpServer(
        name: "TestServer",
        version: "1.0",
        maxConcurrentRequests: 100
    );
    
    server.RegisterTool(new FastExecutingTool());
    await server.StartAsync();
    
    var client = new McpClient("http://localhost:5000");
    
    // Act
    var tasks = new List<Task<McpResponse>>();
    for (int i = 0; i < 1000; i++)
    {
        tasks.Add(client.ExecuteToolAsync("fastTool", new { iteration = i }));
    }
    
    var results = await Task.WhenAll(tasks);
    
    // Assert
    Assert.Equal(1000, results.Length);
    Assert.All(results, r => Assert.NotNull(r));
}

2. ストレステスト

極端な負荷下でシステムをテストしてください:

@Test
public void testServerUnderStress() {
    int maxUsers = 1000;
    int rampUpTimeSeconds = 60;
    int testDurationSeconds = 300;
    
    // Set up JMeter for stress testing
    StandardJMeterEngine jmeter = new StandardJMeterEngine();
    
    // Configure JMeter test plan
    HashTree testPlanTree = new HashTree();
    
    // Create test plan, thread group, samplers, etc.
    TestPlan testPlan = new TestPlan("MCP Server Stress Test");
    testPlanTree.add(testPlan);
    
    ThreadGroup threadGroup = new ThreadGroup();
    threadGroup.setNumThreads(maxUsers);
    threadGroup.setRampUp(rampUpTimeSeconds);
    threadGroup.setScheduler(true);
    threadGroup.setDuration(testDurationSeconds);
    
    testPlanTree.add(threadGroup);
    
    // Add HTTP sampler for tool execution
    HTTPSampler toolExecutionSampler = new HTTPSampler();
    toolExecutionSampler.setDomain("localhost");
    toolExecutionSampler.setPort(5000);
    toolExecutionSampler.setPath("/mcp/execute");
    toolExecutionSampler.setMethod("POST");
    toolExecutionSampler.addArgument("toolName", "calculator");
    toolExecutionSampler.addArgument("parameters", "{\"operation\":\"add\",\"a\":5,\"b\":7}");
    
    threadGroup.add(toolExecutionSampler);
    
    // Add listeners
    SummaryReport summaryReport = new SummaryReport();
    threadGroup.add(summaryReport);
    
    // Run test
    jmeter.configure(testPlanTree);
    jmeter.run();
    
    // Validate results
    assertEquals(0, summaryReport.getErrorCount());
    assertTrue(summaryReport.getAverage() < 200); // Average response time < 200ms
    assertTrue(summaryReport.getPercentile(90.0) < 500); // 90th percentile < 500ms
}

3. モニタリングとプロファイリング

長期的なパフォーマンス分析のためにモニタリングを設定してください:

# Configure monitoring for an MCP server
def configure_monitoring(server):
    # Set up Prometheus metrics
    prometheus_metrics = {
        "request_count": Counter("mcp_requests_total", "Total MCP requests"),
        "request_latency": Histogram(
            "mcp_request_duration_seconds", 
            "Request duration in seconds",
            buckets=[0.01, 0.05, 0.1, 0.5, 1.0, 2.5, 5.0, 10.0]
        ),
        "tool_execution_count": Counter(
            "mcp_tool_executions_total", 
            "Tool execution count",
            labelnames=["tool_name"]
        ),
        "tool_execution_latency": Histogram(
            "mcp_tool_duration_seconds", 
            "Tool execution duration in seconds",
            labelnames=["tool_name"],
            buckets=[0.01, 0.05, 0.1, 0.5, 1.0, 2.5, 5.0, 10.0]
        ),
        "tool_errors": Counter(
            "mcp_tool_errors_total",
            "Tool execution errors",
            labelnames=["tool_name", "error_type"]
        )
    }
    
    # Add middleware for timing and recording metrics
    server.add_middleware(PrometheusMiddleware(prometheus_metrics))
    
    # Expose metrics endpoint
    @server.router.get("/metrics")
    async def metrics():
        return generate_latest()
    
    return server

MCPワークフロー設計パターン

よく設計されたMCPワークフローは効率性、信頼性、保守性を向上させます。以下は従うべき主要なパターンです:

1. ツールチェーンパターン

複数のツールを連続して接続し、各ツールの出力を次のツールの入力として使用します:

# Python Chain of Tools implementation
class ChainWorkflow:
    def __init__(self, tools_chain):
        self.tools_chain = tools_chain  # List of tool names to execute in sequence
    
    async def execute(self, mcp_client, initial_input):
        current_result = initial_input
        all_results = {"input": initial_input}
        
        for tool_name in self.tools_chain:
            # Execute each tool in the chain, passing previous result
            response = await mcp_client.execute_tool(tool_name, current_result)
            
            # Store result and use as input for next tool
            all_results[tool_name] = response.result
            current_result = response.result
        
        return {
            "final_result": current_result,
            "all_results": all_results
        }

# Example usage
data_processing_chain = ChainWorkflow([
    "dataFetch",
    "dataCleaner",
    "dataAnalyzer",
    "dataVisualizer"
])

result = await data_processing_chain.execute(
    mcp_client,
    {"source": "sales_database", "table": "transactions"}
)

2. ディスパッチャーパターン

入力に基づいて専門ツールにディスパッチする中央ツールを使用します:

public class ContentDispatcherTool : IMcpTool
{
    private readonly IMcpClient _mcpClient;
    
    public ContentDispatcherTool(IMcpClient mcpClient)
    {
        _mcpClient = mcpClient;
    }
    
    public string Name => "contentProcessor";
    public string Description => "Processes content of various types";
    
    public object GetSchema()
    {
        return new {
            type = "object",
            properties = new {
                content = new { type = "string" },
                contentType = new { 
                    type = "string",
                    enum = new[] { "text", "html", "markdown", "csv", "code" }
                },
                operation = new { 
                    type = "string",
                    enum = new[] { "summarize", "analyze", "extract", "convert" }
                }
            },
            required = new[] { "content", "contentType", "operation" }
        };
    }
    
    public async Task<ToolResponse> ExecuteAsync(ToolRequest request)
    {
        var content = request.Parameters.GetProperty("content").GetString();
        var contentType = request.Parameters.GetProperty("contentType").GetString();
        var operation = request.Parameters.GetProperty("operation").GetString();
        
        // Determine which specialized tool to use
        string targetTool = DetermineTargetTool(contentType, operation);
        
        // Forward to the specialized tool
        var specializedResponse = await _mcpClient.ExecuteToolAsync(
            targetTool,
            new { content, options = GetOptionsForTool(targetTool, operation) }
        );
        
        return new ToolResponse { Result = specializedResponse.Result };
    }
    
    private string DetermineTargetTool(string contentType, string operation)
    {
        return (contentType, operation) switch
        {
            ("text", "summarize") => "textSummarizer",
            ("text", "analyze") => "textAnalyzer",
            ("html", _) => "htmlProcessor",
            ("markdown", _) => "markdownProcessor",
            ("csv", _) => "csvProcessor",
            ("code", _) => "codeAnalyzer",
            _ => throw new ToolExecutionException($"No tool available for {contentType}/{operation}")
        };
    }
    
    private object GetOptionsForTool(string toolName, string operation)
    {
        // Return appropriate options for each specialized tool
        return toolName switch
        {
            "textSummarizer" => new { length = "medium" },
            "htmlProcessor" => new { cleanUp = true, operation },
            // Options for other tools...
            _ => new { }
        };
    }
}

3. 並列処理パターン

効率性のために複数のツールを同時に実行します:

public class ParallelDataProcessingWorkflow {
    private final McpClient mcpClient;
    
    public ParallelDataProcessingWorkflow(McpClient mcpClient) {
        this.mcpClient = mcpClient;
    }
    
    public WorkflowResult execute(String datasetId) {
        // Step 1: Fetch dataset metadata (synchronous)
        ToolResponse metadataResponse = mcpClient.executeTool("datasetMetadata", 
            Map.of("datasetId", datasetId));
        
        // Step 2: Launch multiple analyses in parallel
        CompletableFuture<ToolResponse> statisticalAnalysis = CompletableFuture.supplyAsync(() ->
            mcpClient.executeTool("statisticalAnalysis", Map.of(
                "datasetId", datasetId,
                "type", "comprehensive"
            ))
        );
        
        CompletableFuture<ToolResponse> correlationAnalysis = CompletableFuture.supplyAsync(() ->
            mcpClient.executeTool("correlationAnalysis", Map.of(
                "datasetId", datasetId,
                "method", "pearson"
            ))
        );
        
        CompletableFuture<ToolResponse> outlierDetection = CompletableFuture.supplyAsync(() ->
            mcpClient.executeTool("outlierDetection", Map.of(
                "datasetId", datasetId,
                "sensitivity", "medium"
            ))
        );
        
        // Wait for all parallel tasks to complete
        CompletableFuture<Void> allAnalyses = CompletableFuture.allOf(
            statisticalAnalysis, correlationAnalysis, outlierDetection
        );
        
        allAnalyses.join();  // Wait for completion
        
        // Step 3: Combine results
        Map<String, Object> combinedResults = new HashMap<>();
        combinedResults.put("metadata", metadataResponse.getResult());
        combinedResults.put("statistics", statisticalAnalysis.join().getResult());
        combinedResults.put("correlations", correlationAnalysis.join().getResult());
        combinedResults.put("outliers", outlierDetection.join().getResult());
        
        // Step 4: Generate summary report
        ToolResponse summaryResponse = mcpClient.executeTool("reportGenerator", 
            Map.of("analysisResults", combinedResults));
        
        // Return complete workflow result
        WorkflowResult result = new WorkflowResult();
        result.setDatasetId(datasetId);
        result.setAnalysisResults(combinedResults);
        result.setSummaryReport(summaryResponse.getResult());
        
        return result;
    }
}

4. エラー回復パターン

ツールの失敗に対する優雅なフォールバックを実装します:

class ResilientWorkflow:
    def __init__(self, mcp_client):
        self.client = mcp_client
    
    async def execute_with_fallback(self, primary_tool, fallback_tool, parameters):
        try:
            # Try primary tool first
            response = await self.client.execute_tool(primary_tool, parameters)
            return {
                "result": response.result,
                "source": "primary",
                "tool": primary_tool
            }
        except ToolExecutionException as e:
            # Log the failure
            logging.warning(f"Primary tool '{primary_tool}' failed: {str(e)}")
            
            # Fall back to secondary tool
            try:
                # Might need to transform parameters for fallback tool
                fallback_params = self._adapt_parameters(parameters, primary_tool, fallback_tool)
                
                response = await self.client.execute_tool(fallback_tool, fallback_params)
                return {
                    "result": response.result,
                    "source": "fallback",
                    "tool": fallback_tool,
                    "primaryError": str(e)
                }
            except ToolExecutionException as fallback_error:
                # Both tools failed
                logging.error(f"Both primary and fallback tools failed. Fallback error: {str(fallback_error)}")
                raise WorkflowExecutionException(
                    f"Workflow failed: primary error: {str(e)}; fallback error: {str(fallback_error)}"
                )
    
    def _adapt_parameters(self, params, from_tool, to_tool):
        """Adapt parameters between different tools if needed"""
        # This implementation would depend on the specific tools
        # For this example, we'll just return the original parameters
        return params

# Example usage
async def get_weather(workflow, location):
    return await workflow.execute_with_fallback(
        "premiumWeatherService",  # Primary (paid) weather API
        "basicWeatherService",    # Fallback (free) weather API
        {"location": location}
    )

5. ワークフロー構成パターン

単純なワークフローを構成して複雑なワークフローを構築します:

public class CompositeWorkflow : IWorkflow
{
    private readonly List<IWorkflow> _workflows;
    
    public CompositeWorkflow(IEnumerable<IWorkflow> workflows)
    {
        _workflows = new List<IWorkflow>(workflows);
    }
    
    public async Task<WorkflowResult> ExecuteAsync(WorkflowContext context)
    {
        var results = new Dictionary<string, object>();
        
        foreach (var workflow in _workflows)
        {
            var workflowResult = await workflow.ExecuteAsync(context);
            
            // Store each workflow's result
            results[workflow.Name] = workflowResult;
            
            // Update context with the result for the next workflow
            context = context.WithResult(workflow.Name, workflowResult);
        }
        
        return new WorkflowResult(results);
    }
    
    public string Name => "CompositeWorkflow";
    public string Description => "Executes multiple workflows in sequence";
}

// Example usage
var documentWorkflow = new CompositeWorkflow(new IWorkflow[] {
    new DocumentFetchWorkflow(),
    new DocumentProcessingWorkflow(),
    new InsightGenerationWorkflow(),
    new ReportGenerationWorkflow()
});

var result = await documentWorkflow.ExecuteAsync(new WorkflowContext {
    Parameters = new { documentId = "12345" }
});

MCPサーバーのテスト:ベストプラクティスとトップヒント

概要

テストは、信頼性が高く高品質なMCPサーバーを開発するための重要な側面です。このガイドでは、単体テストから統合テスト、エンドツーエンドの検証まで、開発ライフサイクル全体を通じてMCPサーバーをテストするための包括的なベストプラクティスとヒントを提供します。

MCPサーバーのテストが重要な理由

MCPサーバーは、AIモデルとクライアントアプリケーションの間の重要なミドルウェアとして機能します。徹底的なテストにより以下を保証します:

  • 本番環境での信頼性
  • リクエストとレスポンスの正確な処理
  • MCP仕様の適切な実装
  • 失敗やエッジケースに対する耐性
  • 様々な負荷下での一貫したパフォーマンス

MCPサーバーの単体テスト

単体テスト(基礎)

単体テストは、MCPサーバーの個々のコンポーネントを単独で検証します。

テストすべき項目

  1. リソースハンドラー: 各リソースハンドラーのロジックを独立してテスト
  2. ツール実装: 様々な入力でツールの動作を検証
  3. プロンプトテンプレート: プロンプトテンプレートが正しくレンダリングされることを確認
  4. スキーマ検証: パラメータ検証ロジックをテスト
  5. エラーハンドリング: 無効な入力に対するエラーレスポンスを検証

単体テストのベストプラクティス

// Example unit test for a calculator tool in C#
[Fact]
public async Task CalculatorTool_Add_ReturnsCorrectSum()
{
    // Arrange
    var calculator = new CalculatorTool();
    var parameters = new Dictionary<string, object>
    {
        ["operation"] = "add",
        ["a"] = 5,
        ["b"] = 7
    };
    
    // Act
    var response = await calculator.ExecuteAsync(parameters);
    var result = JsonSerializer.Deserialize<CalculationResult>(response.Content[0].ToString());
    
    // Assert
    Assert.Equal(12, result.Value);
}
# Example unit test for a calculator tool in Python
def test_calculator_tool_add():
    # Arrange
    calculator = CalculatorTool()
    parameters = {
        "operation": "add",
        "a": 5,
        "b": 7
    }
    
    # Act
    response = calculator.execute(parameters)
    result = json.loads(response.content[0].text)
    
    # Assert
    assert result["value"] == 12

統合テスト(中間層)

統合テストは、MCPサーバーのコンポーネント間の相互作用を検証します。

テストすべき項目

  1. サーバー初期化: 様々な構成でサーバーの起動をテスト
  2. ルート登録: すべてのエンドポイントが正しく登録されていることを確認
  3. リクエスト処理: リクエストとレスポンスの完全なサイクルをテスト
  4. エラー伝播: コンポーネント間でエラーが適切に処理されることを確認
  5. 認証と認可: セキュリティメカニズムをテスト

統合テストのベストプラクティス

// Example integration test for MCP server in C#
[Fact]
public async Task Server_ProcessToolRequest_ReturnsValidResponse()
{
    // Arrange
    var server = new McpServer();
    server.RegisterTool(new CalculatorTool());
    await server.StartAsync();
    
    var request = new McpRequest
    {
        Tool = "calculator",
        Parameters = new Dictionary<string, object>
        {
            ["operation"] = "multiply",
            ["a"] = 6,
            ["b"] = 7
        }
    };
    
    // Act
    var response = await server.ProcessRequestAsync(request);
    
    // Assert
    Assert.NotNull(response);
    Assert.Equal(McpStatusCodes.Success, response.StatusCode);
    // Additional assertions for response content
    
    // Cleanup
    await server.StopAsync();
}

エンドツーエンドテスト(最上層)

エンドツーエンドテストは、クライアントからサーバーまでの完全なシステム動作を検証します。

テストすべき項目

  1. クライアント-サーバー通信: 完全なリクエスト-レスポンスサ
  2. パフォーマンスの基準: パフォーマンスのベンチマークを維持し、リグレッションを検出する
  3. セキュリティスキャン: パイプラインの一部としてセキュリティテストを自動化する

CIパイプラインの例 (GitHub Actions)

name: MCP Server Tests

on:
  push:
    branches: [ main ]
  pull_request:
    branches: [ main ]

jobs:
  test:
    runs-on: ubuntu-latest
    
    steps:
    - uses: actions/checkout@v2
    
    - name: Set up Runtime
      uses: actions/setup-dotnet@v1
      with:
        dotnet-version: '8.0.x'
    
    - name: Restore dependencies
      run: dotnet restore
    
    - name: Build
      run: dotnet build --no-restore
    
    - name: Unit Tests
      run: dotnet test --no-build --filter Category=Unit
    
    - name: Integration Tests
      run: dotnet test --no-build --filter Category=Integration
      
    - name: Performance Tests
      run: dotnet run --project tests/PerformanceTests/PerformanceTests.csproj

MCP仕様への準拠テスト

サーバーがMCP仕様を正しく実装しているか確認します。

主な準拠エリア

  1. APIエンドポイント: 必須エンドポイント (/resources, /tools など) をテストする
  2. リクエスト/レスポンス形式: スキーマ準拠を検証する
  3. エラーコード: 各種シナリオにおける正しいステータスコードを確認する
  4. コンテンツタイプ: 異なるコンテンツタイプの処理をテストする
  5. 認証フロー: 仕様に準拠した認証メカニズムを確認する

準拠テストスイート

[Fact]
public async Task Server_ResourceEndpoint_ReturnsCorrectSchema()
{
    // Arrange
    var client = new HttpClient();
    client.DefaultRequestHeaders.Add("Authorization", "Bearer test-token");
    
    // Act
    var response = await client.GetAsync("http://localhost:5000/api/resources");
    var content = await response.Content.ReadAsStringAsync();
    var resources = JsonSerializer.Deserialize<ResourceList>(content);
    
    // Assert
    Assert.Equal(HttpStatusCode.OK, response.StatusCode);
    Assert.NotNull(resources);
    Assert.All(resources.Resources, resource => 
    {
        Assert.NotNull(resource.Id);
        Assert.NotNull(resource.Type);
        // Additional schema validation
    });
}

効果的なMCPサーバーテストのためのトップ10のヒント

  1. ツール定義を個別にテストする: スキーマ定義をツールロジックとは独立して検証する
  2. パラメータ化されたテストを使用する: 様々な入力(エッジケースを含む)でツールをテストする
  3. エラーレスポンスを確認する: すべての可能なエラー条件に対して適切なエラーハンドリングを検証する
  4. 認可ロジックをテストする: 異なるユーザーロールに対する適切なアクセス制御を確認する
  5. テストカバレッジを監視する: 重要なコードパスの高いカバレッジを目指す
  6. ストリーミングレスポンスをテストする: ストリーミングコンテンツの適切な処理を確認する
  7. ネットワーク問題をシミュレートする: ネットワークが不安定な状況での挙動をテストする
  8. リソース制限をテストする: クォータやレート制限に達した場合の挙動を確認する
  9. リグレッションテストを自動化する: コード変更ごとに実行されるスイートを構築する
  10. テストケースを文書化する: テストシナリオの明確なドキュメントを維持する

よくあるテストの落とし穴

  • ハッピーパステストへの過度の依存: エラーケースを徹底的にテストすることを忘れない
  • パフォーマンステストの無視: 本番環境に影響を与える前にボトルネックを特定する
  • 孤立したテストのみを実施: ユニットテスト、統合テスト、E2Eテストを組み合わせる
  • APIカバレッジの不完全さ: すべてのエンドポイントと機能をテストする
  • 一貫性のないテスト環境: コンテナを使用して一貫したテスト環境を確保する

結論

包括的なテスト戦略は、信頼性が高く高品質なMCPサーバーを開発するために不可欠です。このガイドで紹介したベストプラクティスやヒントを実践することで、MCP実装が最高水準の品質、信頼性、パフォーマンスを満たすことを保証できます。

主なポイント

  1. ツール設計: 単一責任の原則に従い、依存性注入を使用し、再利用可能な設計を目指す
  2. スキーマ設計: 明確で適切に文書化されたスキーマを作成し、適切なバリデーション制約を設ける
  3. エラーハンドリング: 優雅なエラーハンドリング、構造化されたエラーレスポンス、リトライロジックを実装する
  4. パフォーマンス: キャッシング、非同期処理、リソーススロットリングを活用する
  5. セキュリティ: 徹底的な入力バリデーション、認可チェック、機密データの適切な取り扱いを行う
  6. テスト: 包括的なユニットテスト、統合テスト、エンドツーエンドテストを作成する
  7. ワークフローパターン: チェーン、ディスパッチャー、並列処理などの確立されたパターンを適用する

演習

以下のシナリオに適したMCPツールとワークフローを設計してください:

  1. 複数の形式のドキュメント (PDF, DOCX, TXT) を受け入れる
  2. ドキュメントからテキストと主要な情報を抽出する
  3. ドキュメントを種類や内容で分類する
  4. 各ドキュメントの要約を生成する

このシナリオに最適なツールスキーマ、エラーハンドリング、ワークフローパターンを実装してください。また、この実装をどのようにテストするかを考慮してください。

リソース

  1. 最新情報を得るために Azure AI Foundry Discord Community のMCPコミュニティに参加する
  2. オープンソースの MCPプロジェクト に貢献する
  3. 自身の組織のAIイニシアチブにMCPの原則を適用する
  4. 業界向けに特化したMCP実装を探求する
  5. マルチモーダル統合やエンタープライズアプリケーション統合など、特定のMCPトピックに関する高度なコースを受講する
  6. ハンズオンラボ を通じて学んだ原則を使い、自分自身のMCPツールやワークフローを構築してみる

次: ベストプラクティス ケーススタディ

免責事項:
この文書はAI翻訳サービスCo-op Translatorを使用して翻訳されています。正確性を追求しておりますが、自動翻訳には誤りや不正確な部分が含まれる可能性があります。元の言語で記載された文書が正式な情報源とみなされるべきです。重要な情報については、専門の人間による翻訳を推奨します。この翻訳の使用に起因する誤解や誤解について、当社は責任を負いません。