如何使用模型上下文协议工具（预览版）

创建项目客户端

创建包含用于连接到 AI 项目和其他资源的终结点的客户端对象。

var projectEndpoint = System.Environment.GetEnvironmentVariable("PROJECT_ENDPOINT");
var modelDeploymentName = System.Environment.GetEnvironmentVariable("MODEL_DEPLOYMENT_NAME");
var mcpServerUrl = System.Environment.GetEnvironmentVariable("MCP_SERVER_URL");
var mcpServerLabel = System.Environment.GetEnvironmentVariable("MCP_SERVER_LABEL");

PersistentAgentsClient agentClient = new(projectEndpoint, new DefaultAzureCredential());

创建 MCP 工具定义

创建 MCP 工具定义并配置允许的工具。

// Create MCP tool definition
MCPToolDefinition mcpTool = new(mcpServerLabel, mcpServerUrl);

// Configure allowed tools (optional)
string searchApiCode = "search_azure_rest_api_code";
mcpTool.AllowedTools.Add(searchApiCode);

在 MCPToolDefinition 代理初始化期间使用。

PersistentAgent agent = agentClient.Administration.CreateAgent(
   model: modelDeploymentName,
   name: "my-mcp-agent",
   instructions: "You are a helpful agent that can use MCP tools to assist users. Use the available MCP tools to answer questions and perform tasks.",
   tools: [mcpTool]);

创建线程并添加消息

创建线程，添加包含代理问题的消息，并使用 MCP 工具资源启动运行。

PersistentAgentThread thread = agentClient.Threads.CreateThread();

// Create message to thread
PersistentThreadMessage message = agentClient.Messages.CreateMessage(
    thread.Id,
    MessageRole.User,
    "Please summarize the Azure REST API specifications Readme");

MCPToolResource mcpToolResource = new(mcpServerLabel);
mcpToolResource.UpdateHeader("SuperSecret", "123456");
ToolResources toolResources = mcpToolResource.ToToolResources();

// Run the agent with MCP tool resources
ThreadRun run = agentClient.Runs.CreateRun(thread, agent, toolResources);

// Handle run execution and tool approvals
while (run.Status == RunStatus.Queued || run.Status == RunStatus.InProgress || run.Status == RunStatus.RequiresAction)
{
    Thread.Sleep(TimeSpan.FromMilliseconds(1000));
    run = agentClient.Runs.GetRun(thread.Id, run.Id);

    if (run.Status == RunStatus.RequiresAction && run.RequiredAction is SubmitToolApprovalAction toolApprovalAction)
    {
        var toolApprovals = new List<ToolApproval>();
        foreach (var toolCall in toolApprovalAction.SubmitToolApproval.ToolCalls)
        {
            if (toolCall is RequiredMcpToolCall mcpToolCall)
            {
                Console.WriteLine($"Approving MCP tool call: {mcpToolCall.Name}, Arguments: {mcpToolCall.Arguments}");
                toolApprovals.Add(new ToolApproval(mcpToolCall.Id, approve: true)
                {
                    Headers = { ["SuperSecret"] = "123456" }
                });
            }
        }

        if (toolApprovals.Count > 0)
        {
            run = agentClient.Runs.SubmitToolOutputsToRun(thread.Id, run.Id, toolApprovals: toolApprovals);
        }
    }
}

打印消息

Pageable<PersistentThreadMessage> messages = agentClient.Messages.GetMessages(
    threadId: thread.Id,
    order: ListSortOrder.Ascending
);

foreach (PersistentThreadMessage threadMessage in messages)
{
    Console.Write($"{threadMessage.CreatedAt:yyyy-MM-dd HH:mm:ss} - {threadMessage.Role,10}: ");
    foreach (MessageContent contentItem in threadMessage.ContentItems)
    {
        if (contentItem is MessageTextContent textItem)
        {
            Console.Write(textItem.Text);
        }
        else if (contentItem is MessageImageFileContent imageFileItem)
        {
            Console.Write($"<image from ID: {imageFileItem.FileId}>");
        }
        Console.WriteLine();
    }
}

可选：删除代理

完成使用代理后，可以通过以下方法删除它：

agentClient.Threads.DeleteThread(threadId: thread.Id);
agentClient.Administration.DeleteAgent(agentId: agent.Id);

使用 MCP 工具创建代理

以下代码示例首先设置必要的导入、获取相关的 MCP 服务器配置，并初始化 AI Project 客户端。 然后，它会创建代理，将消息添加到线程，并运行代理。

# Import necessary libraries

import os, time
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential
from azure.ai.agents.models import (
    ListSortOrder,
    McpTool,
    RequiredMcpToolCall,
    RunStepActivityDetails,
    SubmitToolApprovalAction,
    ToolApproval,
)

# Get MCP server configuration from environment variables
mcp_server_url = os.environ.get("MCP_SERVER_URL", "https://gitmcp.io/Azure/azure-rest-api-specs")
mcp_server_label = os.environ.get("MCP_SERVER_LABEL", "github")

project_client = AIProjectClient(
    endpoint=os.environ["PROJECT_ENDPOINT"],
    credential=DefaultAzureCredential(),
)
# Initialize agent MCP tool
mcp_tool = McpTool(
    server_label=mcp_server_label,
    server_url=mcp_server_url,
    allowed_tools=[],  # Optional: specify allowed tools
)

# You can also add or remove allowed tools dynamically
search_api_code = "search_azure_rest_api_code"
mcp_tool.allow_tool(search_api_code)
print(f"Allowed tools: {mcp_tool.allowed_tools}")

# Create agent with MCP tool and process agent run
with project_client:
    agents_client = project_client.agents

    # Create a new agent.
    # NOTE: To reuse existing agent, fetch it with get_agent(agent_id)
    agent = agents_client.create_agent(
        model=os.environ["MODEL_DEPLOYMENT_NAME"],
        name="my-mcp-agent",
        instructions="You are a helpful agent that can use MCP tools to assist users. Use the available MCP tools to answer questions and perform tasks.",
        tools=mcp_tool.definitions,
    )

    print(f"Created agent, ID: {agent.id}")
    print(f"MCP Server: {mcp_tool.server_label} at {mcp_tool.server_url}")

    # Create thread for communication
    thread = agents_client.threads.create()
    print(f"Created thread, ID: {thread.id}")

    # Create message to thread
    message = agents_client.messages.create(
        thread_id=thread.id,
        role="user",
        content="Please summarize the Azure REST API specifications Readme",
    )
    print(f"Created message, ID: {message.id}")
    # Create and process agent run in thread with MCP tools
    mcp_tool.update_headers("SuperSecret", "123456")
    # mcp_tool.set_approval_mode("never")  # Uncomment to disable approval requirement
    run = agents_client.runs.create(thread_id=thread.id, agent_id=agent.id, tool_resources=mcp_tool.resources)
    print(f"Created run, ID: {run.id}")

    while run.status in ["queued", "in_progress", "requires_action"]:
        time.sleep(1)
        run = agents_client.runs.get(thread_id=thread.id, run_id=run.id)

        if run.status == "requires_action" and isinstance(run.required_action, SubmitToolApprovalAction):
            tool_calls = run.required_action.submit_tool_approval.tool_calls
            if not tool_calls:
                print("No tool calls provided - cancelling run")
                agents_client.runs.cancel(thread_id=thread.id, run_id=run.id)
                break

            tool_approvals = []
            for tool_call in tool_calls:
                if isinstance(tool_call, RequiredMcpToolCall):
                    try:
                        print(f"Approving tool call: {tool_call}")
                        tool_approvals.append(
                            ToolApproval(
                                tool_call_id=tool_call.id,
                                approve=True,
                                headers=mcp_tool.headers,
                            )
                        )
                    except Exception as e:
                        print(f"Error approving tool_call {tool_call.id}: {e}")

            print(f"tool_approvals: {tool_approvals}")
            if tool_approvals:
                agents_client.runs.submit_tool_outputs(
                    thread_id=thread.id, run_id=run.id, tool_approvals=tool_approvals
                )

        print(f"Current run status: {run.status}")

    print(f"Run completed with status: {run.status}")
    if run.status == "failed":
        print(f"Run failed: {run.last_error}")

    # Display run steps and tool calls
    run_steps = agents_client.run_steps.list(thread_id=thread.id, run_id=run.id)

    # Loop through each step
    for step in run_steps:
        print(f"Step {step['id']} status: {step['status']}")

        # Check if there are tool calls in the step details
        step_details = step.get("step_details", {})
        tool_calls = step_details.get("tool_calls", [])

        if tool_calls:
            print("  MCP Tool calls:")
            for call in tool_calls:
                print(f"    Tool Call ID: {call.get('id')}")
                print(f"    Type: {call.get('type')}")

        if isinstance(step_details, RunStepActivityDetails):
            for activity in step_details.activities:
                for function_name, function_definition in activity.tools.items():
                    print(
                        f'  The function {function_name} with description "{function_definition.description}" will be called.:'
                    )
                    if len(function_definition.parameters) > 0:
                        print("  Function parameters:")
                        for argument, func_argument in function_definition.parameters.properties.items():
                            print(f"      {argument}")
                            print(f"      Type: {func_argument.type}")
                            print(f"      Description: {func_argument.description}")
                    else:
                        print("This function has no parameters")

        print()  # add an extra newline between steps

    # Fetch and log all messages
    messages = agents_client.messages.list(thread_id=thread.id, order=ListSortOrder.ASCENDING)
    print("\nConversation:")
    print("-" * 50)
    for msg in messages:
        if msg.text_messages:
            last_text = msg.text_messages[-1]
            print(f"{msg.role.upper()}: {last_text.text.value}")
            print("-" * 50)

    # Example of dynamic tool management
    print(f"\nDemonstrating dynamic tool management:")
    print(f"Current allowed tools: {mcp_tool.allowed_tools}")

    # Remove a tool
    try:
        mcp_tool.disallow_tool(search_api_code)
        print(f"After removing {search_api_code}: {mcp_tool.allowed_tools}")
    except ValueError as e:
        print(f"Error removing tool: {e}")

    # Clean-up and delete the agent once the run is finished.
    # NOTE: Comment out this line if you plan to reuse the agent later.
    agents_client.delete_agent(agent.id)
    print("Deleted agent")

后续步骤

• 工具概述
• Python SDK 示例

按照 REST API 快速入门 为环境变量 AGENT_TOKEN、AZURE_AI_FOUNDRY_PROJECT_ENDPOINT 和 API_VERSION 设置正确的值。

创建启用了 MCP 工具的代理

若要使 MCP 工具可用于你的代理，请通过服务器终结点、服务器标签等初始化该工具：

curl --request POST \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/assistants?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d "{
    "instructions": "You are a customer support chatbot. Use the tools provided and your knowledge base to best respond to customer queries.",
    "tools": [
          {
              "type": "mcp",
              "server_label": "<unique name for your MCP server>",
              "server_url": "<your MCP server URL>",
              "allowed_tools": ["<tool_name>"], # optional
          }
      ],
  "name": "my-assistant",
  "model": "gpt-4o",
}"

创建线程

curl --request POST \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d ''

将用户问题添加到线程

curl --request POST \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads/thread_abc123/messages?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
      "role": "user",
      "content": "<user input related to the MCP server you connect>"
    }'

创建运行并检查输出

创建一个执行过程以传递工具的标头。 观察模型是否使用“使用必应搜索提供事实依据”工具来响应用户的问题。

require_approval 参数是可选的。 支持的值有：

• always：开发人员需要为每个呼叫提供审批。 如果未提供值，则此值为默认值。
• never：无需审批。
• {"never":[<tool_name_1>, <tool_name_2>]}：提供不需要审批的工具列表。
• {"always":[<tool_name_1>, <tool_name_2>]}：提供需要审批的工具列表。

curl --request POST \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads/thread_abc123/runs?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
      "assistant_id": "<agent_id>",
      "tool_resources": {
          "mcp": [
            {
                "server_label": "<the same unique name you provided during agent creation>",
    "require_approval": "always" #always by default
                "headers": {
                    "Authorization": "Bearer <token>",
                }

            }
          ]
      },
    }'

获取运行状态

curl --request GET \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads/thread_abc123/runs/run_abc123?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN"

如果模型尝试调用 MCP 服务器中需要审批的工具，你会得到一个运行状态为 requires_action 的运行：

{
  "id": "run_123",
  "object": "thread.run",
  ...
  "status": "requires_action",
  ...
  "required_action": {
    "type": "submit_tool_approval",
    "submit_tool_approval": {
      "tool_calls": [
        {
          "id": "call_123",
          "type": "mcp",
          "arguments": "{...}",
          "name": "<tool_name>",
          "server_label": "<server_label_you_provided>"
        }
      ]
    }
  },
  ...
  "tools": [
    {
      "type": "mcp",
      "server_label": "<server_label_you_provided>",
      "server_url": "<server_url_you_provided>",
      "allowed_tools": null
    }
  ],
 ...
}

仔细查看要传递的工具和参数，以便你可以做出明智的审批决策。

提交您的批准

如果决定批准，请将 approve 参数设置为 true，使用 id 参数值应用于之前的工具调用：

curl --request POST \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads/thread_abc123/runs/run_abc123/submit_tool_outputs?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
  "tool_approvals": [
        {
            "tool_call_id": "call_abc123",
            "approve": true,
            "headers": {
            }
        }
    ]
}

检索代理响应

curl --request GET \
  --url $AZURE_AI_FOUNDRY_PROJECT_ENDPOINT/threads/thread_abc123/messages?api-version=$API_VERSION \
  -H "Authorization: Bearer $AGENT_TOKEN"