MCP Development Best Practices
(Click the image above to view video of this lesson)
Overview
This lesson focuses on advanced best practices for developing, testing, and deploying MCP servers and features in production environments. As MCP ecosystems grow in complexity and importance, following established patterns ensures reliability, maintainability, and interoperability. This lesson consolidates practical wisdom gained from real-world MCP implementations to guide you in creating robust, efficient servers with effective resources, prompts, and tools.
Learning Objectives
By the end of this lesson, you will be able to:
- Apply industry best practices in MCP server and feature design
- Create comprehensive testing strategies for MCP servers
- Design efficient, reusable workflow patterns for complex MCP applications
- Implement proper error handling, logging, and observability in MCP servers
- Optimize MCP implementations for performance, security, and maintainability
MCP Core Principles
Before diving into specific implementation practices, it’s important to understand the core principles that guide effective MCP development:
Standardized Communication: MCP uses JSON-RPC 2.0 as its foundation, providing a consistent format for requests, responses, and error handling across all implementations.
User-Centric Design: Always prioritize user consent, control, and transparency in your MCP implementations.
Security First: Implement robust security measures including authentication, authorization, validation, and rate limiting.
Modular Architecture: Design your MCP servers with a modular approach, where each tool and resource has a clear, focused purpose.
Stateful Connections: Leverage MCP’s ability to maintain state across multiple requests for more coherent and context-aware interactions.
Official MCP Best Practices
The following best practices are derived from the official Model Context Protocol documentation:
Security Best Practices
User Consent and Control: Always require explicit user consent before accessing data or performing operations. Provide clear control over what data is shared and which actions are authorized.
Data Privacy: Only expose user data with explicit consent and protect it with appropriate access controls. Safeguard against unauthorized data transmission.
Tool Safety: Require explicit user consent before invoking any tool. Ensure users understand each tool’s functionality and enforce robust security boundaries.
Tool Permission Control: Configure which tools a model is allowed to use during a session, ensuring only explicitly authorized tools are accessible.
Authentication: Require proper authentication before granting access to tools, resources, or sensitive operations using API keys, OAuth tokens, or other secure authentication methods.
Parameter Validation: Enforce validation for all tool invocations to prevent malformed or malicious input from reaching tool implementations.
Rate Limiting: Implement rate limiting to prevent abuse and ensure fair usage of server resources.
Implementation Best Practices
Capability Negotiation: During connection setup, exchange information about supported features, protocol versions, available tools, and resources.
Tool Design: Create focused tools that do one thing well, rather than monolithic tools that handle multiple concerns.
Error Handling: Implement standardized error messages and codes to help diagnose issues, handle failures gracefully, and provide actionable feedback.
Logging: Configure structured logs for auditing, debugging, and monitoring protocol interactions.
Progress Tracking: For long-running operations, report progress updates to enable responsive user interfaces.
Request Cancellation: Allow clients to cancel in-flight requests that are no longer needed or taking too long.
Additional References
For the most up-to-date information on MCP best practices, refer to:
Practical Implementation Examples
Tool Design Best Practices
1. Single Responsibility Principle
Each MCP tool should have a clear, focused purpose. Rather than creating monolithic tools that attempt to handle multiple concerns, develop specialized tools that excel at specific tasks.
// 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. Consistent Error Handling
Implement robust error handling with informative error messages and appropriate recovery mechanisms.
# 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. Parameter Validation
Always validate parameters thoroughly to prevent malformed or malicious input.
// 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
// ...
}
}
Security Implementation Examples
1. Authentication and Authorization
// 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. Rate Limiting
// 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
// ...
}
}
Testing Best Practices
1. Unit Testing MCP Tools
Always test your tools in isolation, mocking external dependencies:
// 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. Integration Testing
Test the complete flow from client requests to server responses:
# 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()
Performance Optimization
1. Caching Strategies
Implement appropriate caching to reduce latency and resource usage:
// 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. Dependency Injection and Testability
Design tools to receive their dependencies through constructor injection, making them testable and configurable:
// 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. Composable Tools
Design tools that can be composed together to create more complex workflows:
# 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
Schema Design Best Practices
The schema is the contract between the model and your tool. Well-designed schemas lead to better tool usability.
1. Clear Parameter Descriptions
Always include descriptive information for each parameter:
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. Validation Constraints
Include validation constraints to prevent invalid inputs:
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. Consistent Return Structures
Maintain consistency in your response structures to make it easier for models to interpret results:
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
}
Error Handling
Robust error handling is crucial for MCP tools to maintain reliability.
1. Graceful Error Handling
Handle errors at appropriate levels and provide informative messages:
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. Structured Error Responses
Return structured error information when possible:
@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. Retry Logic
Implement appropriate retry logic for transient failures:
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)}")
Performance Optimization
1. Caching
Implement caching for expensive operations:
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. Asynchronous Processing
Use asynchronous programming patterns for I/O-bound operations:
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. Resource Throttling
Implement resource throttling to prevent overload:
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
Security Best Practices
1. Input Validation
Always validate input parameters thoroughly:
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. Authorization Checks
Implement proper authorization checks:
@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. Sensitive Data Handling
Handle sensitive data carefully:
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
Testing Best Practices for MCP Tools
Comprehensive testing ensures that MCP tools function correctly, handle edge cases, and integrate properly with the rest of the system.
Unit Testing
1. Test Each Tool in Isolation
Create focused tests for each tool’s functionality:
[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. Schema Validation Testing
Test that schemas are valid and properly enforce constraints:
@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. Error Handling Tests
Create specific tests for error conditions:
@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
Integration Testing
1. Tool Chain Testing
Test tools working together in expected combinations:
[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 Server Testing
Test the MCP server with full tool registration and execution:
@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. End-to-End Testing
Test complete workflows from model prompt to tool execution:
@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"
Performance Testing
1. Load Testing
Test how many concurrent requests your MCP server can handle:
[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. Stress Testing
Test the system under extreme load:
@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. Monitoring and Profiling
Set up monitoring for long-term performance analysis:
# 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 Workflow Design Patterns
Well-designed MCP workflows improve efficiency, reliability, and maintainability. Here are key patterns to follow:
1. Chain of Tools Pattern
Connect multiple tools in a sequence where each tool’s output becomes the input for the next:
# 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. Dispatcher Pattern
Use a central tool that dispatches to specialized tools based on input:
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. Parallel Processing Pattern
Execute multiple tools simultaneously for efficiency:
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. Error Recovery Pattern
Implement graceful fallbacks for tool failures:
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. Workflow Composition Pattern
Build complex workflows by composing simpler ones:
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" }
});
Testing MCP Servers: Best Practices and Top Tips
Overview
Testing is a critical aspect of developing reliable, high-quality MCP servers. This guide provides comprehensive best practices and tips for testing your MCP servers throughout the development lifecycle, from unit tests to integration tests and end-to-end validation.
Why Testing Matters for MCP Servers
MCP servers serve as crucial middleware between AI models and client applications. Thorough testing ensures:
- Reliability in production environments
- Accurate handling of requests and responses
- Proper implementation of MCP specifications
- Resilience against failures and edge cases
- Consistent performance under various loads
Unit Testing for MCP Servers
Unit Testing (Foundation)
Unit tests verify individual components of your MCP server in isolation.
What to Test
- Resource Handlers: Test each resource handler’s logic independently
- Tool Implementations: Verify tool behavior with various inputs
- Prompt Templates: Ensure prompt templates render correctly
- Schema Validation: Test parameter validation logic
- Error Handling: Verify error responses for invalid inputs
Best Practices for Unit Testing
// 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
Integration Testing (Middle Layer)
Integration tests verify interactions between components of your MCP server.
What to Test
- Server Initialization: Test server startup with various configurations
- Route Registration: Verify all endpoints are correctly registered
- Request Processing: Test the full request-response cycle
- Error Propagation: Ensure errors are properly handled across components
- Authentication & Authorization: Test security mechanisms
Best Practices for Integration Testing
// 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();
}
End-to-End Testing (Top Layer)
End-to-end tests verify the complete system behavior from client to server.
What to Test
- Client-Server Communication: Test complete request-response cycles
- Real Client SDKs: Test with actual client implementations
- Performance Under Load: Verify behavior with multiple concurrent requests
- Error Recovery: Test system recovery from failures
- Long-Running Operations: Verify handling of streaming and long operations
Best Practices for E2E Testing
// Example E2E test with a client in TypeScript
describe('MCP Server E2E Tests', () => {
let client: McpClient;
beforeAll(async () => {
// Start server in test environment
await startTestServer();
client = new McpClient('http://localhost:5000');
});
afterAll(async () => {
await stopTestServer();
});
test('Client can invoke calculator tool and get correct result', async () => {
// Act
const response = await client.invokeToolAsync('calculator', {
operation: 'divide',
a: 20,
b: 4
});
// Assert
expect(response.statusCode).toBe(200);
expect(response.content[0].text).toContain('5');
});
});
Mocking Strategies for MCP Testing
Mocking is essential for isolating components during testing.
Components to Mock
- External AI Models: Mock model responses for predictable testing
- External Services: Mock API dependencies (databases, third-party services)
- Authentication Services: Mock identity providers
- Resource Providers: Mock expensive resource handlers
Example: Mocking an AI Model Response
// C# example with Moq
var mockModel = new Mock<ILanguageModel>();
mockModel
.Setup(m => m.GenerateResponseAsync(
It.IsAny<string>(),
It.IsAny<McpRequestContext>()))
.ReturnsAsync(new ModelResponse {
Text = "Mocked model response",
FinishReason = FinishReason.Completed
});
var server = new McpServer(modelClient: mockModel.Object);
# Python example with unittest.mock
@patch('mcp_server.models.OpenAIModel')
def test_with_mock_model(mock_model):
# Configure mock
mock_model.return_value.generate_response.return_value = {
"text": "Mocked model response",
"finish_reason": "completed"
}
# Use mock in test
server = McpServer(model_client=mock_model)
# Continue with test
Performance Testing
Performance testing is crucial for production MCP servers.
What to Measure
- Latency: Response time for requests
- Throughput: Requests handled per second
- Resource Utilization: CPU, memory, network usage
- Concurrency Handling: Behavior under parallel requests
- Scaling Characteristics: Performance as load increases
Tools for Performance Testing
- k6: Open-source load testing tool
- JMeter: Comprehensive performance testing
- Locust: Python-based load testing
- Azure Load Testing: Cloud-based performance testing
Example: Basic Load Test with k6
// k6 script for load testing MCP server
import http from 'k6/http';
import { check, sleep } from 'k6';
export const options = {
vus: 10, // 10 virtual users
duration: '30s',
};
export default function () {
const payload = JSON.stringify({
tool: 'calculator',
parameters: {
operation: 'add',
a: Math.floor(Math.random() * 100),
b: Math.floor(Math.random() * 100)
}
});
const params = {
headers: {
'Content-Type': 'application/json',
'Authorization': 'Bearer test-token'
},
};
const res = http.post('http://localhost:5000/api/tools/invoke', payload, params);
check(res, {
'status is 200': (r) => r.status === 200,
'response time < 500ms': (r) => r.timings.duration < 500,
});
sleep(1);
}
Test Automation for MCP Servers
Automating your tests ensures consistent quality and faster feedback loops.
CI/CD Integration
- Run Unit Tests on Pull Requests: Ensure code changes don’t break existing functionality
- Integration Tests in Staging: Run integration tests in pre-production environments
- Performance Baselines: Maintain performance benchmarks to catch regressions
- Security Scans: Automate security testing as part of the pipeline
Example CI Pipeline (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
Testing for Compliance with MCP Specification
Verify your server correctly implements the MCP specification.
Key Compliance Areas
- API Endpoints: Test required endpoints (/resources, /tools, etc.)
- Request/Response Format: Validate schema compliance
- Error Codes: Verify correct status codes for various scenarios
- Content Types: Test handling of different content types
- Authentication Flow: Verify spec-compliant auth mechanisms
Compliance Test Suite
[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
});
}
Top 10 Tips for Effective MCP Server Testing
- Test Tool Definitions Separately: Verify schema definitions independently from tool logic
- Use Parameterized Tests: Test tools with a variety of inputs, including edge cases
- Check Error Responses: Verify proper error handling for all possible error conditions
- Test Authorization Logic: Ensure proper access control for different user roles
- Monitor Test Coverage: Aim for high coverage of critical path code
- Test Streaming Responses: Verify proper handling of streaming content
- Simulate Network Issues: Test behavior under poor network conditions
- Test Resource Limits: Verify behavior when reaching quotas or rate limits
- Automate Regression Tests: Build a suite that runs on every code change
- Document Test Cases: Maintain clear documentation of test scenarios
Common Testing Pitfalls
- Over-reliance on happy path testing: Make sure to test error cases thoroughly
- Ignoring performance testing: Identify bottlenecks before they affect production
- Testing in isolation only: Combine unit, integration, and E2E tests
- Incomplete API coverage: Ensure all endpoints and features are tested
- Inconsistent test environments: Use containers to ensure consistent test environments
Conclusion
A comprehensive testing strategy is essential for developing reliable, high-quality MCP servers. By implementing the best practices and tips outlined in this guide, you can ensure your MCP implementations meet the highest standards of quality, reliability, and performance.
Key Takeaways
- Tool Design: Follow single responsibility principle, use dependency injection, and design for composability
- Schema Design: Create clear, well-documented schemas with proper validation constraints
- Error Handling: Implement graceful error handling, structured error responses, and retry logic
- Performance: Use caching, asynchronous processing, and resource throttling
- Security: Apply thorough input validation, authorization checks, and sensitive data handling
- Testing: Create comprehensive unit, integration, and end-to-end tests
- Workflow Patterns: Apply established patterns like chains, dispatchers, and parallel processing
Exercise
Design an MCP tool and workflow for a document processing system that:
- Accepts documents in multiple formats (PDF, DOCX, TXT)
- Extracts text and key information from the documents
- Classifies documents by type and content
- Generates a summary of each document
Implement the tool schemas, error handling, and a workflow pattern that best suits this scenario. Consider how you would test this implementation.
Resources
- Join the MCP community on the Azure AI Foundry Discord Community to stay updated on the latest developments
- Contribute to open-source MCP projects
- Apply MCP principles in your own organization’s AI initiatives
- Explore specialized MCP implementations for your industry.
- Consider taking advanced courses on specific MCP topics, such as multi-modal integration or enterprise application integration.
- Experiment with building your own MCP tools and workflows using the principles learned through the Hands on Lab
Next: Best Practices case studies