Spring AI MCP Client与MCP Server交互异常排查求助

我正在实现一个简单的MCP Server与Client，Server仅提供创建文件的单一工具。Server端通过Postman测试MCP请求正常，但Client端已确认和Server建立连接，调用Azure OpenAI大模型时却无法触发Server端的工具执行。日志显示Client已接收Server返回的工具元数据，求排查问题原因。

Server端实现 #

工具服务代码 #

@Slf4j
@Service
public class FileSystemService {

    @Tool(name = "Create File", description = "Create a file with the provided fileName on the file system")
    public String createFile(String fileName) {

       log.info("Request to create a file: {}", fileName);

       Path path = Paths.get(fileName);

       try {
           Files.createFile(path);
           log.info("File created: {}", path.toAbsolutePath());
       } catch (IOException e) {
           log.error("Error creating file: {}", e.getMessage());
       }

       return "File [" + fileName + "] created successfully";
   }
}

Server工具暴露配置 #

@Configuration
public class ToolConfiguration {

    @Bean
    public ToolCallbackProvider tools(FileSystemService fileSystemService) {

        return MethodToolCallbackProvider.builder()
                .toolObjects(fileSystemService)
                .build();
    }
}

Server配置文件（application.yaml） #

# Spring configurations
spring:
  application:
    name: mcp-filesystem-server

  # Spring AI configurations
  ai:
    mcp:
      server:
        enabled: true
        stdio: false
        name: mcp-filesystem-server
        version: 0.0.1
        resource-change-notification: true
        tool-change-notification: true
        prompt-change-notification: true
        sse-endpoint: /api/v1/sse
        sse-message-endpoint: /api/v1/mcp
        type: async

Client端实现 #

Client配置文件（application.yaml） #

spring:
  application:
    name: mcp-client-x

  # Spring AI configurations
  ai:

    # Azure OpenAI configurations
    azure:
      openai:
        chat:
          options:
            deployment-name: gpt-4o-mini
        endpoint: https://ai-foundry-deployment.openai.azure.com/
        api-key: bsbfldsablf
        embeddings:
          options:
            deployment-name: text-embedding-ada-002

    # MCP Client to Server Connection Configuration
    mcp:
      client:
        enabled: true
        name: mcp-client-x
        version: 0.0.1
        initialized: true
        request-timeout: 20s
        root-change-notification: true
        type: async
        toolcallback.enabled: true
        sse:
          connections:
            server1:
              url: http://localhost:8018
              sse-endpoint: /api/v1/sse

ChatClient配置代码 #

@Configuration
public class ModelConfiguration {

    @Bean
    public ChatClient chatClient(AzureOpenAiChatModel azureOpenAiChatModel,
                                 List<ToolCallback> mcpAsyncToolCallbacks) {

        return ChatClient.builder(azureOpenAiChatModel)
                .defaultToolCallbacks(mcpAsyncToolCallbacks)
                .build();
    }
}

LLM调用代码 #

@Slf4j
@Service
@RequiredArgsConstructor
public class ChatServiceAzureOpenAi implements ChatService {

    private final ChatClient chatClient;

    @Override
    public Flux<String> process(UserPrompt userPrompt) {

        log.info("User prompt: {}", userPrompt);

        var SYSTEM_PROMPT = """
                    You are a helpful assistant.
                    Your task is to perform the task based on the user's prompt while making use of the available tools.
                    Message: {message}
                """;

        PromptTemplate promptTemplate = PromptTemplate.builder()
                .template(SYSTEM_PROMPT)
                .build();

        Prompt prompt = promptTemplate.create(Map.of("message", userPrompt.getMessage()));

        return chatClient
                .prompt(prompt)
                .stream()
                .content();
    }
}

关键日志信息 #

2025-08-05T09:54:26.443+05:30 
DEBUG 92639 --- [mcp-client-x] [ctor-http-nio-2] io.modelcontextprotocol.spec.McpSchema   : 
Received JSON message: 
{"jsonrpc":"2.0","id":"8e7ffaaf-1",
"result":{"tools": 
[{"name":"Create File",
"description":"Create a file with the provided fileName on the file system",
"inputSchema": 
{"type":"object","properties":{"arg0":{"type":"string"}},"required":["arg0"],"additionalProperties":false}}]}}

排查方向 #

• 工具参数命名不匹配：日志显示工具输入参数名为arg0，但工具方法参数是fileName，LLM无法正确关联参数。建议给工具方法参数添加注解指定明确名称：

  public String createFile(@Tool.Argument(name = "fileName") String fileName)
  

• 系统提示词不够明确：当前提示词未明确告知LLM调用工具的规则，优化提示词明确要求调用Create File工具并传入fileName参数：

  var SYSTEM_PROMPT = """
              You are a helpful assistant.
              When the user asks to create a file, you must use the `Create File` tool with the `fileName` parameter set to the requested file name.
              Perform the task based on the user's prompt using available tools.
              Message: {message}
          """;
  

• 验证工具回调注入：在ModelConfiguration中添加日志，确认mcpAsyncToolCallbacks是否正确加载：

  @Bean
  public ChatClient chatClient(AzureOpenAiChatModel azureOpenAiChatModel,
                               List<ToolCallback> mcpAsyncToolCallbacks) {
      log.info("Loaded {} MCP tool callbacks", mcpAsyncToolCallbacks.size());
      return ChatClient.builder(azureOpenAiChatModel)
              .defaultToolCallbacks(mcpAsyncToolCallbacks)
              .build();
  }
  

• 检查Azure OpenAI部署配置：确认gpt-4o-mini部署已启用函数调用（Function Calling）功能。
• 开启MCP客户端DEBUG日志：查看工具调用请求是否正常发送到Server端，排查异步通信是否存在问题。

内容的提问来源于stack exchange，提问作者Gagan