Skip to content

Latest commit

 

History

History
553 lines (438 loc) · 14 KB

java-mcp-server.instructions.md

File metadata and controls

553 lines (438 loc) · 14 KB
description applyTo
Best practices and patterns for building Model Context Protocol (MCP) servers in Java using the official MCP Java SDK with reactive streams and Spring integration.
**/*.java, **/pom.xml, **/build.gradle, **/build.gradle.kts

Java MCP Server Development Guidelines

When building MCP servers in Java, follow these best practices and patterns using the official Java SDK.

Dependencies

Add the MCP Java SDK to your Maven project:

<dependencies>
    <dependency>
        <groupId>io.modelcontextprotocol.sdk</groupId>
        <artifactId>mcp</artifactId>
        <version>0.14.1</version>
    </dependency>
</dependencies>

Or for Gradle:

dependencies {
    implementation("io.modelcontextprotocol.sdk:mcp:0.14.1")
}

Server Setup

Create an MCP server using the builder pattern:

import io.mcp.server.McpServer;
import io.mcp.server.McpServerBuilder;
import io.mcp.server.transport.StdioServerTransport;

McpServer server = McpServerBuilder.builder()
    .serverInfo("my-server", "1.0.0")
    .capabilities(capabilities -> capabilities
        .tools(true)
        .resources(true)
        .prompts(true))
    .build();

// Start with stdio transport
StdioServerTransport transport = new StdioServerTransport();
server.start(transport).subscribe();

Adding Tools

Register tool handlers with the server:

import io.mcp.server.tool.Tool;
import io.mcp.server.tool.ToolHandler;
import reactor.core.publisher.Mono;

// Define a tool
Tool searchTool = Tool.builder()
    .name("search")
    .description("Search for information")
    .inputSchema(JsonSchema.object()
        .property("query", JsonSchema.string()
            .description("Search query")
            .required(true))
        .property("limit", JsonSchema.integer()
            .description("Maximum results")
            .defaultValue(10)))
    .build();

// Register tool handler
server.addToolHandler("search", (arguments) -> {
    String query = arguments.get("query").asText();
    int limit = arguments.has("limit") 
        ? arguments.get("limit").asInt() 
        : 10;
    
    // Perform search
    List<String> results = performSearch(query, limit);
    
    return Mono.just(ToolResponse.success()
        .addTextContent("Found " + results.size() + " results")
        .build());
});

Adding Resources

Implement resource handlers for data access:

import io.mcp.server.resource.Resource;
import io.mcp.server.resource.ResourceHandler;

// Register resource list handler
server.addResourceListHandler(() -> {
    List<Resource> resources = List.of(
        Resource.builder()
            .name("Data File")
            .uri("resource://data/example.txt")
            .description("Example data file")
            .mimeType("text/plain")
            .build()
    );
    return Mono.just(resources);
});

// Register resource read handler
server.addResourceReadHandler((uri) -> {
    if (uri.equals("resource://data/example.txt")) {
        String content = loadResourceContent(uri);
        return Mono.just(ResourceContent.text(content, uri));
    }
    throw new ResourceNotFoundException(uri);
});

// Register resource subscribe handler
server.addResourceSubscribeHandler((uri) -> {
    subscriptions.add(uri);
    log.info("Client subscribed to {}", uri);
    return Mono.empty();
});

Adding Prompts

Implement prompt handlers for templated conversations:

import io.mcp.server.prompt.Prompt;
import io.mcp.server.prompt.PromptMessage;
import io.mcp.server.prompt.PromptArgument;

// Register prompt list handler
server.addPromptListHandler(() -> {
    List<Prompt> prompts = List.of(
        Prompt.builder()
            .name("analyze")
            .description("Analyze a topic")
            .argument(PromptArgument.builder()
                .name("topic")
                .description("Topic to analyze")
                .required(true)
                .build())
            .argument(PromptArgument.builder()
                .name("depth")
                .description("Analysis depth")
                .required(false)
                .build())
            .build()
    );
    return Mono.just(prompts);
});

// Register prompt get handler
server.addPromptGetHandler((name, arguments) -> {
    if (name.equals("analyze")) {
        String topic = arguments.getOrDefault("topic", "general");
        String depth = arguments.getOrDefault("depth", "basic");
        
        List<PromptMessage> messages = List.of(
            PromptMessage.user("Please analyze this topic: " + topic),
            PromptMessage.assistant("I'll provide a " + depth + " analysis of " + topic)
        );
        
        return Mono.just(PromptResult.builder()
            .description("Analysis of " + topic + " at " + depth + " level")
            .messages(messages)
            .build());
    }
    throw new PromptNotFoundException(name);
});

Reactive Streams Pattern

The Java SDK uses Reactive Streams (Project Reactor) for asynchronous processing:

// Return Mono for single results
server.addToolHandler("process", (args) -> {
    return Mono.fromCallable(() -> {
        String result = expensiveOperation(args);
        return ToolResponse.success()
            .addTextContent(result)
            .build();
    }).subscribeOn(Schedulers.boundedElastic());
});

// Return Flux for streaming results
server.addResourceListHandler(() -> {
    return Flux.fromIterable(getResources())
        .map(r -> Resource.builder()
            .uri(r.getUri())
            .name(r.getName())
            .build())
        .collectList();
});

Synchronous Facade

For blocking use cases, use the synchronous API:

import io.mcp.server.McpSyncServer;

McpSyncServer syncServer = server.toSyncServer();

// Blocking tool handler
syncServer.addToolHandler("greet", (args) -> {
    String name = args.get("name").asText();
    return ToolResponse.success()
        .addTextContent("Hello, " + name + "!")
        .build();
});

Transport Configuration

Stdio Transport

For local subprocess communication:

import io.mcp.server.transport.StdioServerTransport;

StdioServerTransport transport = new StdioServerTransport();
server.start(transport).block();

HTTP Transport (Servlet)

For HTTP-based servers:

import io.mcp.server.transport.ServletServerTransport;
import jakarta.servlet.http.HttpServlet;

public class McpServlet extends HttpServlet {
    private final McpServer server;
    private final ServletServerTransport transport;
    
    public McpServlet() {
        this.server = createMcpServer();
        this.transport = new ServletServerTransport();
    }
    
    @Override
    protected void doPost(HttpServletRequest req, HttpServletResponse resp) {
        transport.handleRequest(server, req, resp).block();
    }
}

Spring Boot Integration

Use the Spring Boot starter for seamless integration:

<dependency>
    <groupId>io.modelcontextprotocol.sdk</groupId>
    <artifactId>mcp-spring-boot-starter</artifactId>
    <version>0.14.1</version>
</dependency>

Configure the server with Spring:

import org.springframework.context.annotation.Configuration;
import io.mcp.spring.McpServerConfigurer;

@Configuration
public class McpConfiguration {
    
    @Bean
    public McpServerConfigurer mcpServerConfigurer() {
        return server -> server
            .serverInfo("spring-server", "1.0.0")
            .capabilities(cap -> cap
                .tools(true)
                .resources(true)
                .prompts(true));
    }
}

Register handlers as Spring beans:

import org.springframework.stereotype.Component;
import io.mcp.spring.ToolHandler;

@Component
public class SearchToolHandler implements ToolHandler {
    
    @Override
    public String getName() {
        return "search";
    }
    
    @Override
    public Tool getTool() {
        return Tool.builder()
            .name("search")
            .description("Search for information")
            .inputSchema(JsonSchema.object()
                .property("query", JsonSchema.string().required(true)))
            .build();
    }
    
    @Override
    public Mono<ToolResponse> handle(JsonNode arguments) {
        String query = arguments.get("query").asText();
        return Mono.just(ToolResponse.success()
            .addTextContent("Search results for: " + query)
            .build());
    }
}

Error Handling

Use proper error handling with MCP exceptions:

server.addToolHandler("risky", (args) -> {
    return Mono.fromCallable(() -> {
        try {
            String result = riskyOperation(args);
            return ToolResponse.success()
                .addTextContent(result)
                .build();
        } catch (ValidationException e) {
            return ToolResponse.error()
                .message("Invalid input: " + e.getMessage())
                .build();
        } catch (Exception e) {
            log.error("Unexpected error", e);
            return ToolResponse.error()
                .message("Internal error occurred")
                .build();
        }
    });
});

JSON Schema Construction

Use the fluent schema builder:

import io.mcp.json.JsonSchema;

JsonSchema schema = JsonSchema.object()
    .property("name", JsonSchema.string()
        .description("User's name")
        .minLength(1)
        .maxLength(100)
        .required(true))
    .property("age", JsonSchema.integer()
        .description("User's age")
        .minimum(0)
        .maximum(150))
    .property("email", JsonSchema.string()
        .description("Email address")
        .format("email")
        .required(true))
    .property("tags", JsonSchema.array()
        .items(JsonSchema.string())
        .uniqueItems(true))
    .additionalProperties(false)
    .build();

Logging and Observability

Use SLF4J for logging:

import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

private static final Logger log = LoggerFactory.getLogger(MyMcpServer.class);

server.addToolHandler("process", (args) -> {
    log.info("Tool called: process, args: {}", args);
    
    return Mono.fromCallable(() -> {
        String result = process(args);
        log.debug("Processing completed successfully");
        return ToolResponse.success()
            .addTextContent(result)
            .build();
    }).doOnError(error -> {
        log.error("Processing failed", error);
    });
});

Propagate context with Reactor:

import reactor.util.context.Context;

server.addToolHandler("traced", (args) -> {
    return Mono.deferContextual(ctx -> {
        String traceId = ctx.get("traceId");
        log.info("Processing with traceId: {}", traceId);
        
        return Mono.just(ToolResponse.success()
            .addTextContent("Processed")
            .build());
    });
});

Testing

Write tests using the synchronous API:

import org.junit.jupiter.api.Test;
import static org.assertj.core.Assertions.assertThat;

class McpServerTest {
    
    @Test
    void testToolHandler() {
        McpServer server = createTestServer();
        McpSyncServer syncServer = server.toSyncServer();
        
        JsonNode args = objectMapper.createObjectNode()
            .put("query", "test");
        
        ToolResponse response = syncServer.callTool("search", args);
        
        assertThat(response.isError()).isFalse();
        assertThat(response.getContent()).hasSize(1);
    }
}

Jackson Integration

The SDK uses Jackson for JSON serialization. Customize as needed:

import com.fasterxml.jackson.databind.ObjectMapper;
import com.fasterxml.jackson.datatype.jsr310.JavaTimeModule;

ObjectMapper mapper = new ObjectMapper();
mapper.registerModule(new JavaTimeModule());

// Use custom mapper with server
McpServer server = McpServerBuilder.builder()
    .objectMapper(mapper)
    .build();

Content Types

Support multiple content types in responses:

import io.mcp.server.content.Content;

server.addToolHandler("multi", (args) -> {
    return Mono.just(ToolResponse.success()
        .addTextContent("Plain text response")
        .addImageContent(imageBytes, "image/png")
        .addResourceContent("resource://data", "application/json", jsonData)
        .build());
});

Server Lifecycle

Properly manage server lifecycle:

import reactor.core.Disposable;

Disposable serverDisposable = server.start(transport).subscribe();

// Graceful shutdown
Runtime.getRuntime().addShutdownHook(new Thread(() -> {
    log.info("Shutting down MCP server");
    serverDisposable.dispose();
    server.stop().block();
}));

Common Patterns

Request Validation

server.addToolHandler("validate", (args) -> {
    if (!args.has("required_field")) {
        return Mono.just(ToolResponse.error()
            .message("Missing required_field")
            .build());
    }
    
    return processRequest(args);
});

Async Operations

server.addToolHandler("async", (args) -> {
    return Mono.fromCallable(() -> callExternalApi(args))
        .timeout(Duration.ofSeconds(30))
        .onErrorResume(TimeoutException.class, e -> 
            Mono.just(ToolResponse.error()
                .message("Operation timed out")
                .build()))
        .subscribeOn(Schedulers.boundedElastic());
});

Resource Caching

private final Map<String, String> cache = new ConcurrentHashMap<>();

server.addResourceReadHandler((uri) -> {
    return Mono.fromCallable(() -> 
        cache.computeIfAbsent(uri, this::loadResource))
        .map(content -> ResourceContent.text(content, uri));
});

Best Practices

  1. Use Reactive Streams for async operations and backpressure
  2. Leverage Spring Boot starter for enterprise applications
  3. Implement proper error handling with specific error messages
  4. Use SLF4J for logging, not System.out
  5. Validate inputs in tool and prompt handlers
  6. Support graceful shutdown with proper resource cleanup
  7. Use bounded elastic scheduler for blocking operations
  8. Propagate context for observability in reactive chains
  9. Test with synchronous API for simplicity
  10. Follow Java naming conventions (camelCase for methods, PascalCase for classes)