聊聊langchain4j的MCP

最新推荐文章于 2025-06-07 14:38:04 发布

hello_ejb3

最新推荐文章于 2025-06-07 14:38:04 发布

阅读量1.5k

点赞数 14

CC 4.0 BY-SA版权

文章标签：人工智能

本文链接：https://blog.youkuaiyun.com/hello_ejb3/article/details/146462924

序

本文主要研究一下langchain4j对Model Context Protocol (MCP) 的支持

MCP

MCP协议规定了两种传输方式：

HTTP：客户端请求一个SSE（Server-Sent Events）通道以从服务器接收事件，然后通过HTTP POST请求发送命令。这种方式适用于需要跨网络通信的场景，通常用于分布式系统或需要高并发的场景。
stdio：客户端可以将MCP服务器作为本地子进程运行，并通过标准输入/输出直接与其通信。这种方式适用于本地集成和命令行工具，适合简单的本地批处理任务。

如果需要让ChatModel或AI service运行由MCP服务器提供的工具，则需要创建一个MCP tool provider

McpToolProvider

McpTransport

McpTransport transport = new StdioMcpTransport.Builder()
    .command(List.of("/usr/bin/npm", "exec", "@modelcontextprotocol/server-everything@0.6.2"))
    .logEvents(true) // only if you want to see the traffic in the log
    .build();

对于HTTP方式则如下：

McpTransport transport = new HttpMcpTransport.Builder()
    .sseUrl("http://localhost:3001/sse")
    .logRequests(true) // if you want to see the traffic in the log
    .logResponses(true)
    .build();

McpClient

McpClient mcpClient = new DefaultMcpClient.Builder()
    .transport(transport)
    .logMessageHandler(new MyLogMessageHandler())
    .build();

基于McpTransport创建McpClient，可以指定logMessageHandler来打印相关信息，可以通过client.listResources()、client.listResourceTemplates()来返回MCP的resources

provider

ToolProvider toolProvider = McpToolProvider.builder()
    .mcpClients(List.of(mcpClient))
    .build();

基于McpClient来创建McpToolProvider

AiService

Bot bot = AiServices.builder(Bot.class)
    .chatLanguageModel(model)
    .toolProvider(toolProvider)
    .build();

AiServices.builder提供了方法来设置toolProvider

示例

public static void main(String[] args) throws Exception {

    ChatLanguageModel model = OpenAiChatModel.builder()
        .apiKey(System.getenv("OPENAI_API_KEY"))
        .modelName("gpt-4o-mini")
        .logRequests(true)
        .logResponses(true)
        .build();

    McpTransport transport = new StdioMcpTransport.Builder()
        .command(List.of("/usr/local/bin/docker", "run", "-e", "GITHUB_PERSONAL_ACCESS_TOKEN", "-i", "mcp/github"))
        .logEvents(true)
        .build();

    McpClient mcpClient = new DefaultMcpClient.Builder()
        .transport(transport)
        .build();

    ToolProvider toolProvider = McpToolProvider.builder()
        .mcpClients(List.of(mcpClient))
        .build();

    Bot bot = AiServices.builder(Bot.class)
        .chatLanguageModel(model)
        .toolProvider(toolProvider)
        .build();

    try {
        String response = bot.chat("Summarize the last 3 commits of the LangChain4j GitHub repository");
        System.out.println("RESPONSE: " + response);
    } finally {
        mcpClient.close();
    }
}

这里使用stdio的方式来连接docker部署的GitHub MCP server，通过chat告诉LLM去总结LangChain4j这个github仓库最近的3个commits

返回结果示例如下：

Here are the summaries of the last three commits in the LangChain4j GitHub repository:

1. **Commit [36951f9](https://github.com/langchain4j/langchain4j/commit/36951f9649c1beacd8b9fc2d910a2e23223e0d93)** (Date: 2025-02-05)
   - **Author:** Dmytro Liubarskyi
   - **Message:** Updated to `upload-pages-artifact@v3`.
   - **Details:** This commit updates the GitHub Action used for uploading pages artifacts to version 3.

2. **Commit [6fcd19f](https://github.com/langchain4j/langchain4j/commit/6fcd19f50c8393729a0878d6125b0bb1967ac055)** (Date: 2025-02-05)
   - **Author:** Dmytro Liubarskyi
   - **Message:** Updated to `checkout@v4`, `deploy-pages@v4`, and `upload-pages-artifact@v4`.
   - **Details:** This commit updates multiple GitHub Actions to their version 4.

3. **Commit [2e74049](https://github.com/langchain4j/langchain4j/commit/2e740495d2aa0f16ef1c05cfcc76f91aef6f6599)** (Date: 2025-02-05)
   - **Author:** Dmytro Liubarskyi
   - **Message:** Updated to `setup-node@v4` and `configure-pages@v4`.
   - **Details:** This commit updates the `setup-node` and `configure-pages` GitHub Actions to version 4.

All commits were made by the same author, Dmytro Liubarskyi, on the same day, focusing on updating various GitHub Actions to newer versions.

源码

ToolProvider

langchain4j/src/main/java/dev/langchain4j/service/tool/ToolProvider.java

@FunctionalInterface
public interface ToolProvider {

    /**
     * Provides tools for the request to the LLM.
     *
     * @param request {@link ToolProviderRequest} contains {@link UserMessage} and chat memory id (see {@link MemoryId}).
     * @return {@link ToolProviderResult} contains tools that should be included in the request to the LLM.
     */
    ToolProviderResult provideTools(ToolProviderRequest request);
}

langchain4j定义了ToolProvider接口，每次调用AI服务时，它都会被调用，并为该次调用提供相应的工具，其provideTools返回ToolProviderResult

McpToolProvider

public class McpToolProvider implements ToolProvider {

    private final List<McpClient> mcpClients;
    private final boolean failIfOneServerFails;
    private static final Logger log = LoggerFactory.getLogger(McpToolProvider.class);

    private McpToolProvider(Builder builder) {
        this.mcpClients = new ArrayList<>(builder.mcpClients);
        this.failIfOneServerFails = Utils.getOrDefault(builder.failIfOneServerFails, false);
    }

    @Override
    public ToolProviderResult provideTools(final ToolProviderRequest request) {
        ToolProviderResult.Builder builder = ToolProviderResult.builder();
        for (McpClient mcpClient : mcpClients) {
            try {
                List<ToolSpecification> toolSpecifications = mcpClient.listTools();
                for (ToolSpecification toolSpecification : toolSpecifications) {
                    builder.add(
                            toolSpecification, (executionRequest, memoryId) -> mcpClient.executeTool(executionRequest));
                }
            } catch (Exception e) {
                if (failIfOneServerFails) {
                    throw new RuntimeException("Failed to retrieve tools from MCP server", e);
                } else {
                    log.warn("Failed to retrieve tools from MCP server", e);
                }
            }
        }
        return builder.build();
    }

    //......
}

McpToolProvider要求构造器传入McpToolProvider.Builder，provideTools会遍历mcpClients，之后遍历mcpClient.listTools()，构建每个tool对应的executor(mcpClient.executeTool(executionRequest))