MCP 生态全景：官方服务器、云托管与 A2A 对比

// 在 Claude Desktop 中挂载官方文件服务器
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/me/projects"
      ]
    }
  }
}

成熟度 = 0.3 × 最近3个月commit数 / 12
       + 0.3 × 周下载量 / 10000
       + 0.2 × 文档完整度（0-1）
       + 0.2 × 安全实践（0-1：是否最小权限/无硬编码密钥）

// Bedrock Agent 的 action group 通过 MCP 连接外部服务器
{
  "agent": {
    "actionGroups": [
      {
        "name": "mcp-tools",
        "type": "MCP",
        "mcpConfig": {
          "url": "https://mcp-internal.example.com/mcp",
          "auth": {
            "type": "IAM",
            "role": "arn:aws:iam::123456789012:role/mcp-gateway-role"
          }
        }
      }
    ]
  }
}

# 使用 Google Cloud 托管的 MCP 端点（示例模式）
from google.cloud.aiplatform import AgentSession
from mcp import ClientSession

# 通过云 SDK 获取短期令牌，再走 Streamable HTTP 传输
transport = StreamableHttpTransport(
    "https://vertex-agent.example.googleapis.com/mcp",
    headers={"Authorization": f"Bearer {gcp_oauth_token()}"},
)
async with ClientSession(transport) as session:
    await session.initialize()
    tools = await session.list_tools()
    result = await session.call_tool(
        tools[0].name, {"query": "检索订单明细"}
    )

{
  "name": "research-agent",
  "description": "负责资料检索与综述",
  "url": "https://agents.example.com/research/a2a",
  "capabilities": {
    "streaming": true,
    "pushNotifications": true
  },
  "skills": [
    { "id": "web-search", "name": "网络搜索" },
    { "id": "summarize", "name": "内容摘要" }
  ]
}

// 客户端向远端 Agent 创建任务
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tasks/send",
  "params": {
    "agentId": "research-agent",
    "task": {
      "id": "task-2026-0001",
      "message": {
        "role": "user",
        "parts": [
          { "text": "调研 2026 年 MCP 生态的最新托管方案，输出 500 字综述" }
        ]
      }
    }
  }
}

// 远端 Agent 的最终响应（含执行状态与产物）
{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "task": {
      "id": "task-2026-0001",
      "status": {
        "state": "completed",
        "message": {
          "role": "agent",
          "parts": [{ "text": "已完成调研……（综述内容省略）" }]
        }
      },
      "artifacts": [
        { "id": "art-1", "name": "survey.md", "kind": "file", "url": "https://…/survey.md" }
      ]
    }
  }
}

# 检查本地 SDK 与规范版本的兼容性
npx @modelcontextprotocol/sdk --version
# 输出形如：1.x.y（遵循 2025-06-18 规范）

// 网关同时挂载 MCP 工具源与 A2A 任务端点
class AgentGateway {
  private mcpTools: Map<string, McpToolHandle> = new Map();
  private a2aTasks: Map<string, A2aTask> = new Map();

  // MCP 侧：接入工具服务器
  async mountMcpServer(name: string, params: StdioServerParameters) {
    const client = await connectMcp(params);
    const { tools } = await client.listTools();
    for (const t of tools) {
      this.mcpTools.set(`${name}__${t.name}`, {
        client, schema: t.inputSchema, description: t.description,
      });
    }
  }

  // A2A 侧：响应远端 Agent 的委托
  async handleA2aTask(agentId: string, message: string): Promise<string> {
    // 把委托任务翻译成内部工具编排
    const plan = await this.planWithTools(message, [...this.mcpTools.keys()]);
    const result = await this.executePlan(plan);
    return result;
  }
}