MCP Workers AI

Name: MCP Workers AI
Author: xtuc

本地服务

github.com/xtuc开发工具699 周前更新

代码仓库

关于 MCP Workers AI

面向 Cloudflare Workers 的 MCP 服务 SDK，可加载其他 MCP 服务提供的工具，并由 Worker 调用。适用于将工具接入 Workers AI 的开发者。

主要功能

01加载 MCP 工具
02调用选定工具
03在 Worker 中运行

适用场景

01在 Worker 中接入 GitLab 工具
02通过自然语言触发工具调用

MCP 相关

作为用于构建 MCP 服务的 SDK，它能加载其他 MCP 服务的工具，并提供调用工具的函数。

配置

本地服务的配置在各客户端中通用，复制后粘贴到客户端的 MCP 配置文件即可。

暂未收录安装方式

我们还没有获取到这个服务的安装包或远程地址，请按项目 README 中的说明安装。

代码仓库

工具

还没有获取到工具列表

本地服务需要在你的电脑上运行后才能读取工具，维护者暂未提供工具列表。

详细介绍

MCP Workers AI

用于 Cloudflare Workers 的 MCP 服务 SDK

用法

安装：

yarn add mcp-workers-ai
# or
npm install -S mcp-workers-ai

加载 MCP 服务工具：

import { loadTools } from "mcp-workers-ai"

const tools = await loadTools([
  import("@modelcontextprotocol/server-gitlab"),
  import("@modelcontextprotocol/server-slack"),
  ...
]);

// Pass `tools` to the LLM inference request.

调用工具：

import { callTool } from "mcp-workers-ai"

// Typically the LLM selects a tool to use.
const selected_tool = {
  arguments: {
    project_id: 'svensauleau/test',
    branch: 'main',
    files: [ ... ],
    commit_message: 'added unit tests'
  },
  name: 'push_files'
};

const res = await callTool(selected_tool)

// Pass `res` back into a LLM inference request.

演示

wrangler 配置：

name = "test"
main = "src/index.ts"

[ai]
binding = "AI"

[vars]
GITLAB_PERSONAL_ACCESS_TOKEN = "glpat-aaaaaaaaaaaaaaaaaaaa"

[alias]
"@modelcontextprotocol/sdk/server/index.js" = "mcp-workers-ai/sdk/server/index.js"
"@modelcontextprotocol/sdk/server/stdio.js" = "mcp-workers-ai/sdk/server/stdio.js"

Worker：

import { loadTools, callTool } from "mcp-workers-ai"

export default {
  async fetch(request: Request, env: any): Promise<Response> {
    // Make sure to set the token before importing the tools
    process.env.GITLAB_PERSONAL_ACCESS_TOKEN = env.GITLAB_PERSONAL_ACCESS_TOKEN;

    const tools = await loadTools([
      import("@modelcontextprotocol/server-gitlab/dist/"),
    ]);

    const prompt = await request.text();

    const response = await env.AI.run(
      "@hf/nousresearch/hermes-2-pro-mistral-7b",
      {
        messages: [{ role: "user", content: prompt }],
        tools,
      },
    );

    if (response.tool_calls && response.tool_calls.length > 0) {
      const selected_tool = response.tool_calls[0];
      const res = await callTool(selected_tool)

      if (res.content.length > 1) {
        throw new Error("too many responses")
      }

      const finalResponse = await env.AI.run(
        "@hf/nousresearch/hermes-2-pro-mistral-7b",
        {
          messages: [
            {
              role: "user",
              content: prompt,
            },
            {
              role: "assistant",
              content: "",
              tool_call: selected_tool.name,
            },
            {
              role: "tool",
              name: selected_tool.name,
              content: res.content[0].text,
            },
          ],
          tools,
        },
      );
      return new Response(finalResponse.response);

    } else {
      return new Response(response.response);
    }
  }
};

调用 AI：

$ curl http://example.com \
  -d "create a file called 'joke.txt' in my svensauleau/test project with your favorite joke on the main branch. Use the commit message 'added unit tests'"

I have successfully added a file called 'joke.txt' with a joke to your project 'svensauleau/test' on the main branch. The commit message used was 'added unit tests'. You can view the commit and the file in your project's repository.

结果：