Organizar a caixa de ferramentas baseada em intenções no Foundry (versão preliminar)

from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import MCPTool, WebSearchTool

# Create Foundry project client
endpoint = "https://<your-foundry-account>.services.ai.azure.com/api/projects/<your-project>"
project = AIProjectClient(
    endpoint=endpoint,
    credential=DefaultAzureCredential(),
)

# Create toolbox version with web search and MCP tools
toolbox_version = project.beta.toolboxes.create_toolbox_version(
    toolbox_name="my-toolbox",
    description="Toolbox with web search and an MCP server",
    tools=[
        WebSearchTool(),
        MCPTool(
            server_label="myserver",
            server_url="https://your-mcp-server.example.com",
            require_approval="never",
            project_connection_id="my-key-auth-connection",
        ),
    ],
)
print(f"Created toolbox: {toolbox_version.name}, version: {toolbox_version.version}")

using Azure.Identity;
using Azure.AI.Projects;

// Create Foundry project client
var projectEndpoint = "https://<your-foundry-account>.services.ai.azure.com/api/projects/<your-project>";
AIProjectClient projectClient = new(new Uri(projectEndpoint), new DefaultAzureCredential());
AgentToolboxes toolboxClient = projectClient.AgentAdministrationClient.GetAgentToolboxes();

ProjectsAgentTool webTool = ProjectsAgentTool.AsProjectTool(
    ResponseTool.CreateWebSearchTool());

ProjectsAgentTool mcpTool = ProjectsAgentTool.AsProjectTool(ResponseTool.CreateMcpTool(
    serverLabel: "myserver",
    serverUri: new Uri("https://your-mcp-server.example.com"),
    toolCallApprovalPolicy: new McpToolCallApprovalPolicy(
        GlobalMcpToolCallApprovalPolicy.NeverRequireApproval
    )
));

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [webTool, mcpTool],
    description: "Toolbox with web search and an MCP server"
);
Console.WriteLine($"Created toolbox: {toolboxVersion.Name}, version: {toolboxVersion.Version}");

POST {project_endpoint}/toolboxes/my-toolbox/versions?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json

{
  "description": "Toolbox with web search and an MCP server",
  "tools": [
    {
      "type": "web_search",
      "description": "Search the web for current information"
    },
    {
      "type": "mcp",
      "server_label": "myserver",
      "server_url": "https://your-mcp-server.example.com",
      "require_approval": "never",
      "project_connection_id": "my-key-auth-connection"
    }
  ]
}

import { DefaultAzureCredential } from "@azure/identity";
import { AIProjectClient } from "@azure/ai-projects";

// Create Foundry project client
const projectEndpoint = "https://<your-foundry-account>.services.ai.azure.com/api/projects/<your-project>";

const project = new AIProjectClient(projectEndpoint, new DefaultAzureCredential());

const toolboxVersion = await project.beta.toolboxes.createVersion(
  "my-toolbox",
  [
    {
      type: "web_search",
      description: "Search the web for current information",
    },
    {
      type: "mcp",
      server_label: "myserver",
      server_url: "https://your-mcp-server.example.com",
      require_approval: "never",
      project_connection_id: "my-key-auth-connection",
    },
  ],
  {
    description: "Toolbox with web search and an MCP server",
  },
);
console.log(`Created toolbox: ${toolboxVersion.name}, version: ${toolboxVersion.version}`);

Use o Foundry Toolkit em Visual Studio Code para criar e publicar uma caixa de ferramentas a partir da vista Tools.

1. Selecione Foundry Toolkit na barra de atividades.
2. Na secção Meus Recursos, expanda o nome> do seuprojeto Ferramentas.
3. Selecione o ícone + Adicionar Caixa de Ferramentas .
4. No separador Construir uma Caixa de Ferramentas Personalizada , introduza o nome e a descrição da caixa de ferramentas, adicione as ferramentas que quiser e depois selecione Publicar.

Publicar uma nova caixa de ferramentas cria a sua primeira versão. Essa versão torna-se automaticamente a versão padrão.

Ao usar azd, declara os recursos da toolbox num agent.yaml ficheiro em vez de chamar o SDK. Defina as suas ferramentas na resources secção e implemente usando azd ai agent init. Para agent.yaml exemplos para cada tipo de ferramenta, veja Configurar ferramentas. Para o fluxo de trabalho completo de implantação, consulte Deploy with azd.

# 1. Create a manifest directory with your agent.yaml + source files
mkdir my-agent/manifest
# Copy agent.yaml, main.py, Dockerfile, requirements.txt into my-agent/manifest/

# 2. Initialize the azd project (note: -m is REQUIRED)
cd my-agent
$PROJECT_ID = "/subscriptions/<sub>/resourceGroups/<rg>/providers/Microsoft.CognitiveServices/accounts/<account>/projects/<project>"
azd ai agent init -m https://github.com/microsoft/hosted-agents-vnext-private-preview/main/samples/python/toolbox/azd/agent.yaml --project-id $PROJECT_ID -e my-env
# Or equivalently: azd ai agent init -m manifest/ --project-id $PROJECT_ID -e my-env
# ↑ If your agent.yaml declares {{ param }} secrets (e.g., github_pat), you will be prompted to enter
#   them interactively HERE — before init completes. This is the only safe time to supply credentials.
# NOTE: Do NOT use --no-prompt here — it skips the prompt and leaves {{ param }} credentials empty (see Troubleshooting: Credentials Empty with --no-prompt)

# 3. CRITICAL post-init fixes (see "Post-Init Checklist" below)
azd env set enableHostedAgentVNext "true" -e my-env
azd env set AZURE_AI_MODEL_DEPLOYMENT_NAME "gpt-4o" -e my-env  # must match the deployment name in azure.yaml

# 4. Provision infrastructure (creates connections via Bicep)
azd provision -e my-env

# 5. Deploy agent (creates toolboxes, container image, agent version)
azd deploy -e my-env

# 6. Invoke the agent (MUST run from the scaffolded project directory)
azd ai agent invoke --new-session "tell me about the latest news in Microsoft Foundry" --timeout 120

Agent.yaml:

kind: hosted
name: toolbox-azd-test
description: LangGraph agent wired for toolbox MCP.
metadata:
  tags:
    - AI Agent Hosting
    - LangGraph

# template: contains the ContainerAgent definition (kind: hosted).
# These fields are used to generate src/<agent>/agent.yaml during init.
template:
  kind: hosted
  protocols:
    - protocol: responses
      version: 1.0.0
  environment_variables:
    # FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_AGENT_TOOLBOX_* are injected
    # automatically by the platform at runtime — do NOT declare them here.
    - name: AZURE_OPENAI_ENDPOINT
      value: ${AZURE_OPENAI_ENDPOINT}
    - name: AZURE_AI_MODEL_DEPLOYMENT_NAME
      value: ${AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-4o}
    - name: TOOLBOX_NAME
      value: ${TOOLBOX_NAME=agent-tools}

# parameters: secret values prompted at init time (or set via azd env).
# azd uppercases the param name to find the env var: github_pat → GITHUB_PAT.
parameters:
  github_pat:
    secret: true
    description: GitHub Personal Access Token (classic ghp_... or fine-grained github_pat_...)

# resources: connections and toolboxes scaffolded into azure.yaml by azd ai agent init.
resources:
  - kind: connection
    name: github-mcp-conn
    target: https://api.githubcopilot.com/mcp
    category: remoteTool
    credentials:
      type: CustomKeys
      keys:
        Authorization: "Bearer {{ github_pat }}"

  - kind: toolbox
    name: agent-tools
    tools:
      - type: web_search
      - type: mcp
        server_label: github
        server_url: https://api.githubcopilot.com/mcp
        project_connection_id: github-mcp-conn

No Foundry Toolkit para Visual Studio Code, copie o endpoint de consumo da caixa de ferramentas da vista Toolboxes.

1. Selecione Foundry Toolkit na barra de atividades.
2. Na secção Meus Recursos, expanda o nome> do seuprojeto Ferramentas.
3. No separador Caixas de Ferramentas , localiza a tua caixa de ferramentas.
4. Na coluna URL do Endpoint, copie o endpoint.

O valor Endpoint URL é o endpoint consumidor da toolbox. Para construir um endpoint específico de cada versão, utilize o padrão de desenvolvimento mostrado na tabela anterior.

Instalar o SDK do cliente MCP:

pip install mcp

Liga-te à caixa de ferramentas e lista ferramentas

import asyncio
from azure.identity import DefaultAzureCredential
from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession

url = "https://<account>.services.ai.azure.com/api/projects/<proj>/toolboxes/<name>/versions/<version>/mcp?api-version=v1"

token = DefaultAzureCredential().get_token("https://ai.azure.com/.default").token
headers = {
    "Authorization": f"Bearer {token}",
    "Foundry-Features": "Toolboxes=V1Preview",
}

async def verify_toolbox():
    async with streamablehttp_client(url, headers=headers) as (read, write, _):
        async with ClientSession(read, write) as session:
            await session.initialize()

            # List available tools
            tools_result = await session.list_tools()
            print(f"Tools found: {len(tools_result.tools)}")
            for tool in tools_result.tools:
                print(f"  - {tool.name}: {(tool.description or '')[:80]}")

            # Call a tool (replace with actual tool name and arguments)
            result = await session.call_tool("<tool_name>", arguments={})
            print(result)

asyncio.run(verify_toolbox())

Use o endpoint específico da versão (/versions/{version}/mcp) para validar uma versão antes de a promover.

1. Inicializar a sessão MCP:

POST {project_endpoint}/toolboxes/{toolbox_name}/versions/{version}/mcp?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json
Foundry-Features: Toolboxes=V1Preview

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}

2. Enviar a notificação inicializada:

POST {project_endpoint}/toolboxes/{toolbox_name}/versions/{version}/mcp?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json
Foundry-Features: Toolboxes=V1Preview

{"jsonrpc":"2.0","method":"notifications/initialized"}

3. Listar as ferramentas disponíveis:

POST {project_endpoint}/toolboxes/{toolbox_name}/versions/{version}/mcp?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json
Foundry-Features: Toolboxes=V1Preview

{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}

4. Chamar uma ferramenta:

POST {project_endpoint}/toolboxes/{toolbox_name}/versions/{version}/mcp?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json
Foundry-Features: Toolboxes=V1Preview

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"<TOOL_NAME>","arguments":{}}}

Instalar o SDK do cliente MCP:

npm install @modelcontextprotocol/sdk

Liga-te à caixa de ferramentas e lista ferramentas

import { DefaultAzureCredential } from "@azure/identity";
import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";
import { Client } from "@modelcontextprotocol/sdk/client/index.js";

const url = "https://<account>.services.ai.azure.com/api/projects/<proj>/toolboxes/<name>/versions/<version>/mcp?api-version=v1";

const credential = new DefaultAzureCredential();
const token = await credential.getToken("https://ai.azure.com/.default");

const transport = new StreamableHTTPClientTransport(
  new URL(url),
  {
    requestInit: {
      headers: {
        Authorization: `Bearer ${token.token}`,
        "Foundry-Features": "Toolboxes=V1Preview",
      },
    },
  },
);

const client = new Client({ name: "test", version: "1.0" });
await client.connect(transport);

// List available tools
const toolsResult = await client.listTools();
console.log(`Tools found: ${toolsResult.tools.length}`);
for (const tool of toolsResult.tools) {
  console.log(`  - ${tool.name}: ${(tool.description || "").slice(0, 80)}`);
}

// Call a tool (replace with actual tool name and arguments)
const result = await client.callTool({ name: "<tool_name>", arguments: {} });
console.log(result);

await client.close();

Use o endpoint do Passo 2 juntamente com um exemplo de agente hospedado e estruturado para validar o carregamento da toolbox no VS Code.

1. No Foundry Toolkit, em Os Meus Recursos>O nome do seu projeto>Ferramentas, localize a caixa de ferramentas que pretende testar.
2. Selecione o modelo de código do andaime.
3. Escolha uma pasta de projeto quando solicitado.
4. Siga o gerado README.md para instalar dependências, configurar variáveis de ambiente e executar a amostra localmente.
5. Usa o Agent Inspector ou executa python main.py para confirmar que as ferramentas da consola carregam e respondem.

Para validação específica da versão antes de promover uma nova versão da toolbox, use o separador Python ou REST API neste passo.

LangGraph

.env ficheiro:

FOUNDRY_PROJECT_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>
FOUNDRY_AGENT_TOOLBOX_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1
TOOLBOX_NAME=agent-tools
FOUNDRY_AGENT_TOOLBOX_FEATURES=Toolboxes=V1Preview
AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-4o

main.py (padrão-chave):

from langchain_azure_ai.tools import AzureAIProjectToolbox

toolbox = AzureAIProjectToolbox(toolbox_name=TOOLBOX_NAME)
tools = await toolbox.get_tools()

Veja a amostra completa para a implementação completa.

Microsoft Agent Framework

Use MCPStreamableHTTPTool a partir do Agent Framework SDK para se ligar diretamente ao endpoint MCP da toolbox.

.env ficheiro:

FOUNDRY_PROJECT_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>
FOUNDRY_AGENT_TOOLBOX_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1
FOUNDRY_AGENT_TOOLBOX_FEATURES=Toolboxes=V1Preview
AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-4o

main.py (padrão-chave):

# Auth: wrap token provider in an httpx.Auth subclass
credential = DefaultAzureCredential()
token_provider = get_bearer_token_provider(credential, "https://ai.azure.com/.default")
http_client = httpx.AsyncClient(
    auth=_ToolboxAuth(token_provider),
    headers={"Foundry-Features": "Toolboxes=V1Preview"},
    timeout=120.0,
)

# Toolbox MCP endpoint (platform-injected at runtime via FOUNDRY_AGENT_TOOLBOX_ENDPOINT)
TOOLBOX_ENDPOINT = "https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1"

# Connect MCPStreamableHTTPTool to the toolbox endpoint
mcp_tool = MCPStreamableHTTPTool(
    name="toolbox",
    url=TOOLBOX_ENDPOINT,
    http_client=http_client,
    load_prompts=False,
)

agent = chat_client.as_agent(
    name="my-toolbox-agent",
    instructions="You are a helpful assistant with access to Foundry toolbox tools.",
    tools=[mcp_tool],
)
ResponsesAgentServerHost().run()

Veja a amostra completa para a implementação completa.

Copilot SDK

Use o GitHub Copilot SDK para construir um agente alimentado por toolbox que liga a invocação de ferramentas do Copilot ao endpoint MCP da Foundry toolbox.

.env ficheiro:

GITHUB_TOKEN=<your-github-token>
FOUNDRY_AGENT_TOOLBOX_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1

agent.py (padrão de chave — ponte MCP):

# 1. Open an MCP session to the toolbox endpoint
bridge = McpBridge(endpoint=TOOLBOX_ENDPOINT, token=_get_toolbox_token())
await bridge.initialize()
mcp_tools = await bridge.list_tools()

# 2. Map MCP tool list to Copilot SDK tool definitions
#    Dots in tool names are replaced with underscores (Copilot SDK requirement)
copilot_tools = [
    {
        "name": t["name"].replace(".", "_"),
        "description": t.get("description", ""),
        "parameters": t.get("inputSchema", {}),
    }
    for t in mcp_tools
]

# 3. Wire tool calls back to the MCP session
async def tool_handler(name: str, arguments: dict) -> str:
    return await bridge.call_tool(name.replace("_", ".", 1), arguments)

# 4. Run the Copilot SDK agent
agent = Agent(
    tools=copilot_tools,
    tool_handler=tool_handler,
    token=os.environ["GITHUB_TOKEN"],
)

Veja a amostra completa para a implementação completa.

Microsoft Agent Framework

Use ResponsesServer a partir do SDK do Agent Framework com um personalizado ToolboxMcpClient para descobrir e invocar ferramentas da caixa de ferramentas através do endpoint MCP.

Variáveis de ambiente:

AZURE_OPENAI_ENDPOINT=https://<account>.services.ai.azure.com
AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-4o
TOOLBOX_MCP_ENDPOINT=https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1

Program.cs (padrão-chave):

using Azure.AI.AgentServer.Responses;
using Azure.AI.AgentServer.Responses.Models;
using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Extensions.DependencyInjection;
using OpenAI.Chat;

// Azure OpenAI endpoint and model deployment
var openAiEndpoint = "https://<account>.services.ai.azure.com";
var deployment = "gpt-4o";  // supports all toolbox tool types

// Toolbox MCP endpoint (platform-injected at runtime via TOOLBOX_MCP_ENDPOINT)
var toolboxEndpoint = "https://<account>.services.ai.azure.com/api/projects/<project>/toolboxes/<toolbox-name>/versions/<version>/mcp?api-version=v1";

// Azure OpenAI client
var credential = new DefaultAzureCredential();
var openAIClient = new AzureOpenAIClient(new Uri(openAiEndpoint), credential);
var chatClient = openAIClient.GetChatClient(deployment);

// Toolbox MCP client — discovers tools via tools/list, calls them via tools/call
var toolboxClient = new ToolboxMcpClient(toolboxEndpoint, credential);

ResponsesServer.Run<ToolboxHandler>(configure: builder =>
{
    builder.Services.AddSingleton(new AgentConfig(chatClient, toolboxClient));
});

ToolboxMcpClient encapsula chamadas JSON-RPC diretamente para o endpoint MCP. ToolboxHandler liga chamadas de ferramentas LLM de volta ao cliente MCP usando um ciclo padrão de chamada de ferramenta. Para a implementação completa de ambas as classes, veja a amostra completa.

Utilize o Foundry Toolkit para criar uma amostra de agente hospedado que já está integrada na sua caixa de ferramentas.

1. Selecione Foundry Toolkit na barra de atividades.
2. Na secção Meus Recursos, expanda o nome> do seuprojeto Ferramentas.
3. No separador Caixas de Ferramentas , localize a caixa de ferramentas que pretende consumir e depois selecione modelo de código de andaime.
4. Na Paleta de Comandos, escolha uma pasta de projeto quando solicitado.
5. Abra o ficheiro gerado README.md e siga os passos de configuração, execução local e implementação da estrutura inicial.

O projeto gerado inclui o ponto de entrada do agente alojado, os ficheiros de implementação e um README.md com os passos exatos de configuração, execução e implementação. O agente com andaimes trata do Foundry-Features: Toolboxes=V1Preview cabeçalho por si.

Se quiser integrar uma caixa de ferramentas num projeto de agente hospedado existente em vez de gerar uma nova amostra, use o endpoint copiado do Step 2 com os padrões Python ou .NET desta secção.

Implementar com azd

Use a CLI de Desenvolvedor Azure (azd) para declarar recursos da caixa de ferramentas diretamente num ficheiro agent.yaml e implante o seu agente com um único comando. Ao usar esta abordagem, não precisa de criar a caixa de ferramentas separadamente através do SDK ou REST. azd Provisiona a caixa de ferramentas, as ligações e a implementação do modelo em conjunto.

Estrutura da pasta:

my-agent/
├── agent.yaml          # Agent, toolbox, and connection declarations
├── main.py             # LangGraph agent
├── requirements.txt    # All dependencies (Azure SDK + PyPI packages)
├── Dockerfile          # Container build

agent.yaml (Exemplo de Pesquisa Web + GitHub MCP):

name: my-toolbox-agent
description: LangGraph agent with Azure AI Foundry toolbox MCP.
metadata:
  tags:
    - AI Agent Hosting
    - LangGraph
template:
  name: my-toolbox-agent
  kind: hosted
  protocols:
    - protocol: responses
      version: 1.0.0
  environment_variables:
    # FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_AGENT_TOOLBOX_* are injected
    # automatically by the platform at runtime — do NOT declare them here.
    - name: AZURE_AI_MODEL_DEPLOYMENT_NAME
      value: ${AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-4o}
    - name: TOOLBOX_NAME
      value: ${TOOLBOX_NAME=agent-tools}
parameters:
  github_pat:
    secret: true
    description: GitHub Personal Access Token for MCP connection
resources:
  - kind: connection
    name: github-mcp-conn
    target: https://api.githubcopilot.com/mcp
    category: RemoteTool
    authType: CustomKeys
    credentials:
      keys:
        Authorization: "Bearer {{ github_pat }}"
  - kind: toolbox
    name: agent-tools
    description: Web search and GitHub MCP tools
    tools:
      - type: web_search
      - type: mcp
        server_label: github
        server_url: https://api.githubcopilot.com/mcp
        project_connection_id: github-mcp-conn

main.py segue o mesmo padrão LangGraph mostrado anteriormente. Ao usar azd, FOUNDRY_AGENT_TOOLBOX_ENDPOINT e TOOLBOX_{toolbox_name}_MCP_ENDPOINT são injetados automaticamente – não é necessária configuração adicional de endpoint no código.

Implementar:

# 1. Place agent.yaml and source files in a manifest directory
mkdir my-agent/manifest
# Copy agent.yaml, main.py, Dockerfile, requirements.txt into my-agent/manifest/

# 2. Initialize the azd project (-m is required)
cd my-agent
PROJECT_ID="/subscriptions/<sub>/resourceGroups/<rg>/providers/Microsoft.CognitiveServices/accounts/<account>/projects/<project>"
azd ai agent init -m manifest/ --project-id $PROJECT_ID -e my-env
# If agent.yaml declares {{ param }} secrets (for example, github_pat), you are prompted
# to enter them interactively here. Do NOT use --no-prompt — it leaves credentials empty.

# 3. Set required environment variables
azd env set enableHostedAgentVNext "true" -e my-env
azd env set AZURE_AI_MODEL_DEPLOYMENT_NAME "gpt-4o" -e my-env

# 4. Provision infrastructure (creates connections via Bicep)
azd provision -e my-env

# 5. Deploy agent (creates toolboxes, container image, agent version)
azd deploy -e my-env

# 6. Invoke the agent
azd ai agent invoke --new-session "Hello, what tools do you have?" --timeout 120

from azure.ai.projects.models import MCPTool

# Set require_approval on an MCP tool
toolbox_version = project.beta.toolboxes.create_toolbox_version(
    toolbox_name="my-toolbox",
    tools=[
        MCPTool(
            server_label="myserver",
            server_url="https://your-mcp-server.example.com",
            require_approval="always",  # "always" | "never"
            project_connection_id="my-connection",
        )
    ],
)

{
  "tools": [
    {
      "type": "mcp",
      "server_label": "myserver",
      "server_url": "https://your-mcp-server.example.com",
      "require_approval": "always",
      "project_connection_id": "my-connection"
    }
  ]
}

ProjectsAgentTool mcpTool = ProjectsAgentTool.AsProjectTool(ResponseTool.CreateMcpTool(
    serverLabel: "myserver",
    serverUri: new Uri("https://your-mcp-server.example.com"),
    toolCallApprovalPolicy: new McpToolCallApprovalPolicy(
        GlobalMcpToolCallApprovalPolicy.AlwaysRequireApproval
    )
));

const tools = [
  {
    type: "mcp",
    server_label: "myserver",
    server_url: "https://your-mcp-server.example.com",
    require_approval: "always",
    project_connection_id: "my-connection",
  },
];

Usa o separador Python, .NET, JavaScript, REST API ou azd para configurar require_approval na definição da tua caixa de ferramentas. O fluxo de trabalho do Foundry Toolkit neste artigo foca-se na criação e consumo da caixa de ferramentas no Visual Studio Code.

resources:
  - kind: toolbox
    name: my-toolbox
    tools:
      - type: mcp
        server_label: myserver
        server_url: https://your-mcp-server.example.com
        require_approval: always
        project_connection_id: my-connection

# Create a new toolbox version
toolbox_version = project.beta.toolboxes.create_toolbox_version(
    toolbox_name="my-toolbox",
    description="Updated tools v2",
    tools=[...],
)
print(f"Created version: {toolbox_version.version}")

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Updated tools v2"
);
Console.WriteLine($"Created version: {toolboxVersion.Version}");

POST {project_endpoint}/toolboxes/my-toolbox/versions?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json

{
  "description": "Updated tools v2",
  "tools": [...]
}

const toolboxVersion = await project.beta.toolboxes.createVersion(
  "my-toolbox",
  [/* tools array */],
  { description: "Updated tools v2" },
);
console.log(`Created version: ${toolboxVersion.version}`);

Use o separador Python, .NET, JavaScript ou API REST para criar uma nova versão da caixa de ferramentas. O fluxo de trabalho do Foundry Toolkit neste artigo foca-se na criação de uma caixa de ferramentas e na estruturação de um agente hospedado que a consome.

Esta operação não é suportada pelo azd. Para criar uma versão da caixa de ferramentas, use o separador Python, .NET, REST ou JavaScript.

# List all toolbox versions
versions = list(project.beta.toolboxes.list_toolbox_versions(toolbox_name="my-toolbox"))
for v in versions:
    print(f"{v.version} — created {v.created_at}")

List<ToolboxVersion> versions = await toolboxClient
    .GetToolboxVersionsAsync("my-toolbox")
    .ToListAsync();
Console.WriteLine($"Found {versions.Count} toolbox version(s).");
foreach (ToolboxVersion v in versions)
{
    Console.WriteLine($"  - {v.Name} ({v.Version})");
}

GET {project_endpoint}/toolboxes/my-toolbox/versions?api-version=v1
Authorization: Bearer {token}

const versions = project.beta.toolboxes.listVersions("my-toolbox");
for await (const v of versions) {
  console.log(`${v.version} — created ${v.createdAt}`);
}

Use o separador Python, .NET, JavaScript ou API REST para listar versões da toolbox.

Esta operação não é suportada pelo azd. Para listar versões das ferramentas, use o separador Python, .NET, API REST, ou JavaScript.

# Get a specific toolbox version
version_obj = project.beta.toolboxes.get_toolbox_version(
    toolbox_name="my-toolbox",
    version="<version_id>",
)

ToolboxVersion versionObj = await toolboxClient.GetToolboxVersionAsync(
    "my-toolbox",
    "<version_id>"
);
Console.WriteLine($"Retrieved toolbox: {versionObj.Name} ({versionObj.Id})");

GET {project_endpoint}/toolboxes/my-toolbox/versions/{version}?api-version=v1
Authorization: Bearer {token}

const versionObj = await project.beta.toolboxes.getVersion(
  "my-toolbox",
  "<version_id>",
);
console.log(`Retrieved version: ${versionObj.version}`);

Use o separador Python, .NET, JavaScript ou API REST para obter uma versão específica da caixa de ferramentas.

Esta operação não é suportada pelo azd. Para obter uma versão específica da caixa de ferramentas, use o separador Python, .NET, REST ou JavaScript.

# Promote a version to default
toolbox = project.beta.toolboxes.update(
    toolbox_name="my-toolbox",
    default_version="<version_id>",
)
print(f"Active version: {toolbox.default_version}")

ToolboxRecord record = await toolboxClient.UpdateToolboxAsync(
    "my-toolbox",
    "<version_id>"
);
Console.WriteLine($"Active version: {record.DefaultVersion}");

PATCH {project_endpoint}/toolboxes/my-toolbox?api-version=v1
Authorization: Bearer {token}
Content-Type: application/json

{
  "default_version": "<version_id>"
}

default_version Não pode estar vazio. Substitua-o por uma nova versão.

const toolbox = await project.beta.toolboxes.update(
  "my-toolbox",
  "<version_id>",
);
console.log(`Active version: ${toolbox.defaultVersion}`);

Utilize o separador Python, .NET, JavaScript ou REST API para promover uma versão da caixa de ferramentas para predefinição.

Esta operação não é suportada pelo azd. Para promover uma versão para o padrão, use o separador Python, .NET, REST ou JavaScript.

# Delete a toolbox version
project.beta.toolboxes.delete_toolbox_version(
    toolbox_name="my-toolbox",
    version="<version_id>",
)

await toolboxClient.DeleteToolboxVersionAsync(
    "my-toolbox",
    "<version_id>"
);

DELETE {project_endpoint}/toolboxes/my-toolbox/versions/{version}?api-version=v1
Authorization: Bearer {token}

await project.beta.toolboxes.deleteVersion(
  "my-toolbox",
  "<version_id>",
);

Use o separador Python, .NET, JavaScript ou REST API para eliminar uma versão da toolbox.

Esta operação não é suportada pelo azd. Para eliminar uma versão de toolbox, use o separador Python, .NET, REST API, ou JavaScript.

Autenticação baseada em teclas:

{
  "description": "my-mcp-toolbox",
  "tools": [
    {
      "type": "mcp",
      "server_label": "myserver",
      "server_url": "https://your-mcp-server.example.com",
      "project_connection_id": "my-mcp-connection"
    }
  ]
}

Sem autenticação (servidor MCP público):

{
  "description": "Public MCP server",
  "tools": [
    {
      "type": "mcp",
      "server_label": "myserver",
      "server_url": "https://your-mcp-server.example.com"
    }
  ]
}

OAuth ou autenticação baseada em identidade:

Para OAuth (conector gerido, registo personalizado de app), identidade de agente ou autenticação do token Entra do utilizador, crie primeiro a ligação apropriada no seu projeto Foundry, depois referencia-a com project_connection_id:

{
  "description": "MCP server with OAuth/identity auth",
  "tools": [
    {
      "type": "mcp",
      "server_label": "myserver",
      "server_url": "https://your-mcp-server.example.com",
      "project_connection_id": "<OAUTH_OR_IDENTITY_CONNECTION_NAME>"
    }
  ]
}

A ligação authType determina o fluxo de autenticação. Os tipos de autenticação de ligação suportados para MCP incluem CustomKeys, OAuth2 (gerida ou personalizada), AgenticIdentity, e UserEntraToken. Consulte o separador azd para exemplos de configuração de ligação para cada tipo de autenticação.

from azure.ai.projects.models import MCPTool

tools = [
    MCPTool(
        server_label="myserver",
        server_url="https://your-mcp-server.example.com",
        project_connection_id="my-mcp-connection",
    )
]

ProjectsAgentTool tool = ProjectsAgentTool.AsProjectTool(ResponseTool.CreateMcpTool(
    serverLabel: "myserver",
    serverUri: new Uri("https://your-mcp-server.example.com")
));

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "my-mcp-toolbox"
);

const tools = [
  {
    type: "mcp",
    server_label: "myserver",
    server_url: "https://your-mcp-server.example.com",
    project_connection_id: "my-mcp-connection",
  },
];

Sem autorização:

resources:
  - kind: toolbox
    name: mcp-tools
    description: Public MCP server tools
    tools:
      - type: mcp
        server_label: myserver
        server_url: https://your-mcp-server.example.com

Autenticação baseada em teclas:

parameters:
  mcp_api_key:
    secret: true
    description: API key for the MCP server
resources:
  - kind: connection
    name: mcp-conn
    target: https://your-mcp-server.example.com
    category: RemoteTool
    authType: CustomKeys
    credentials:
      keys:
        Authorization: "Bearer {{ mcp_api_key }}"
  - kind: toolbox
    name: mcp-tools
    description: MCP server tools with key auth
    tools:
      - type: mcp
        server_label: myserver
        server_url: https://your-mcp-server.example.com
        project_connection_id: mcp-conn

OAuth - conector gerido:

Use este padrão para servidores MCP que suportam o fluxo OAuth gerido pela Foundry. O valor connectorName deve corresponder a um conector administrado disponível no Catálogo de Ferramentas Foundry.

resources:
  - kind: connection
    name: github-oauth-conn
    category: RemoteTool
    authType: OAuth2
    target: https://api.githubcopilot.com/mcp
    connectorName: foundrygithubmcp
  - kind: toolbox
    name: oauth-tools
    description: GitHub OAuth MCP toolbox
    tools:
      - type: mcp
        server_label: github
        project_connection_id: github-oauth-conn

OAuth - registo personalizado de aplicações:

Use este padrão quando levar o seu próprio registo na aplicação OAuth para o servidor MCP.

parameters:
  oauth_client_id:
    secret: true
    description: OAuth client ID
  oauth_client_secret:
    secret: true
    description: OAuth client secret
resources:
  - kind: connection
    name: mcp-oauth-custom-conn
    category: RemoteTool
    authType: OAuth2
    target: https://your-mcp-server.example.com
    authorizationUrl: https://auth.example.com/authorize
    tokenUrl: https://auth.example.com/token
    refreshUrl: https://auth.example.com/token
    scopes: []
    credentials:
      clientID: "{{ oauth_client_id }}"
      clientSecret: "{{ oauth_client_secret }}"
  - kind: toolbox
    name: oauth-custom-tools
    description: MCP toolbox with custom OAuth
    tools:
      - type: mcp
        server_label: myserver
        project_connection_id: mcp-oauth-custom-conn

Identidade do agente (Entra ID):

Use este padrão para servidores MCP que suportam autenticação Microsoft Entra ID. A identidade do agente Foundry autentica-se perante o recurso alvo.

resources:
  - kind: connection
    name: language-mcp
    category: RemoteTool
    authType: AgenticIdentity
    audience: <entra-audience>
    target: https://<resource>.cognitiveservices.azure.com/language/mcp?api-version=2025-11-15-preview
  - kind: toolbox
    name: agent-identity-tools
    description: MCP toolbox with agent identity auth
    tools:
      - type: mcp
        server_label: language
        project_connection_id: language-mcp

Token de Entrada de Utilizador (1P OBO):

Use este padrão para servidores MCP que requerem identidade de utilizador através do fluxo On-Behalf-Of (OBO). A Foundry intermedia o token Entra do utilizador ao servidor MCP.

resources:
  - kind: connection
    name: workiq-mail-conn
    category: RemoteTool
    authType: UserEntraToken
    audience: <entra-app-id>
    target: https://agent365.svc.cloud.microsoft/agents/servers/mcp_MailTools
  - kind: toolbox
    name: workiq-tools
    description: MCP toolbox with user Entra token auth
    tools:
      - type: mcp
        server_label: workiq
        project_connection_id: workiq-mail-conn

{
  "description": "Built-in web search",
  "tools": [
    {
      "type": "web_search",
      "name": "<OPTIONAL_TOOL_NAME>",
      "description": "<Optional description for the model>"
    }
  ]
}

Com uma ligação ao Bing Pesquisa Personalizada:

{
  "description": "Custom Bing Search instance",
  "tools": [
    {
      "type": "web_search",
      "name": "<OPTIONAL_TOOL_NAME>",
      "description": "<Optional description for the model>",
      "web_search": {
        "custom_search_configuration": {
          "project_connection_id": "<BING_CONNECTION_NAME>",
          "instance_name": "<BING_INSTANCE_NAME>"
        }
      }
    }
  ]
}

from azure.ai.projects.models import WebSearchTool

tools = [
    WebSearchTool()
]

ProjectsAgentTool tool = ProjectsAgentTool.AsProjectTool(
    ResponseTool.CreateWebSearchTool()
);

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Built-in web search"
);

const tools = [
  {
    type: "web_search",
    name: "<OPTIONAL_TOOL_NAME>",
    description: "<Optional description for the model>",
  },
];

resources:
  - kind: toolbox
    name: websearch-tools
    description: Web search toolbox
    tools:
      - type: web_search

Integração com a Pesquisa Personalizada do Bing:

parameters:
  bing_api_key:
    secret: true
    description: Bing API key
resources:
  - kind: connection
    name: bing-custom-conn
    category: GroundingWithCustomSearch
    authType: ApiKey
    target: ""
    credentials:
      key: "{{ bing_api_key }}"
    metadata:
      ResourceId: /subscriptions/<sub>/resourceGroups/<rg>/providers/Microsoft.Bing/accounts/<bing-account>
      type: bing_custom_search
  - kind: toolbox
    name: bing-custom-tools
    description: Bing Custom Search toolbox
    tools:
      - type: bing_custom_search
        custom_search_configuration:
          instance_name: your-bing-custom-instance
        project_connection_id: bing-custom-conn

{
  "description": "Azure AI Search over my data",
  "tools": [
    {
      "type": "azure_ai_search",
      "name": "<OPTIONAL_TOOL_NAME>",
      "description": "<Optional description for the model>",
      "azure_ai_search": {
        "indexes": [
          {
            "index_name": "<INDEX_NAME>",
            "project_connection_id": "<CONNECTION_NAME>"
          }
        ]
      }
    }
  ]
}

from azure.ai.projects.models import AzureAISearchTool

tools = [
    AzureAISearchTool(
        index_name="<INDEX_NAME>",
        project_connection_id="<CONNECTION_NAME>",
    )
]

ProjectsAgentTool tool = new AzureAISearchTool(
    new AzureAISearchToolOptions(
        indexes: [
            new AzureAISearchIndexResource(
                indexName: "<INDEX_NAME>",
                projectConnectionId: "<CONNECTION_NAME>"
            )
        ]
    )
);

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Azure AI Search over my data"
);

const tools = [
  {
    type: "azure_ai_search",
    name: "<OPTIONAL_TOOL_NAME>",
    description: "<Optional description for the model>",
    azure_ai_search: {
      indexes: [
        {
          index_name: "<INDEX_NAME>",
          project_connection_id: "<CONNECTION_NAME>",
        },
      ],
    },
  },
];

parameters:
  ai_search_key:
    secret: true
    description: Azure AI Search admin key
resources:
  - kind: connection
    name: aisearch-conn
    category: CognitiveSearch
    authType: ApiKey
    target: https://your-search-service.search.windows.net/
    credentials:
      key: "{{ ai_search_key }}"
  - kind: toolbox
    name: search-tools
    description: Azure AI Search toolbox
    tools:
      - type: azure_ai_search
        index_name: your-index-name
        project_connection_id: aisearch-conn

{
  "description": "Code interpreter for data analysis",
  "tools": [
    {
      "type": "code_interpreter",
      "name": "<OPTIONAL_TOOL_NAME>",
      "description": "<Optional description for the model>",
      "container": {
            "type": "auto",
            "file_ids": ["<FILE_ID>"]
        }
    }
  ]
}

from azure.ai.projects.models import CodeInterpreterTool

tools = [
    CodeInterpreterTool()
]

ProjectsAgentTool tool = ProjectsAgentTool.AsProjectTool(
    ResponseTool.CreateCodeInterpreterTool(
        new CodeInterpreterToolContainer()
    )
);

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Code interpreter for data analysis"
);

const tools = [
  {
    type: "code_interpreter",
    name: "<OPTIONAL_TOOL_NAME>",
    description: "<Optional description for the model>",
    container: {
      type: "auto",
      file_ids: ["<FILE_ID>"],
    },
  },
];

resources:
  - kind: toolbox
    name: codeinterp-tools
    description: Code interpreter toolbox
    tools:
      - type: code_interpreter

{
  "description": "Search over uploaded documents",
  "tools": [
    {
      "type": "file_search",
      "name": "<OPTIONAL_TOOL_NAME>",
      "description": "<Optional description for the model>",
      "file_search": {
        "vector_store_ids": ["<VECTOR_STORE_ID>"]
      }
    }
  ]
}

from azure.ai.projects.models import FileSearchTool

tools = [
    FileSearchTool(
        vector_store_ids=["<VECTOR_STORE_ID>"]
    )
]

ProjectsAgentTool tool = ProjectsAgentTool.AsProjectTool(
    ResponseTool.CreateFileSearchTool(
        vectorStoreIds: ["<VECTOR_STORE_ID>"]
    )
);

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Search over uploaded documents"
);

const tools = [
  {
    type: "file_search",
    name: "<OPTIONAL_TOOL_NAME>",
    description: "<Optional description for the model>",
    file_search: {
      vector_store_ids: ["<VECTOR_STORE_ID>"],
    },
  },
];

resources:
  - kind: toolbox
    name: filesearch-tools
    description: File search toolbox
    tools:
      - type: file_search
        vector_store_ids:
          - ${FILE_SEARCH_VECTOR_STORE_ID}

Defina o ID da loja vetorial antes de implementar:

azd env set FILE_SEARCH_VECTOR_STORE_ID "vs_xxxxxxxxxxxx"

Autorização anónima:

{
  "description": "REST API via OpenAPI spec",
  "tools": [
    {
      "type": "openapi",
      "openapi": {
        "name": "my-api",
        "spec": { "<paste OpenAPI spec object here>" },
        "auth": {
          "type": "anonymous"
        }
      }
    }
  ]
}

Autenticação de ligação do projeto:

Use este padrão quando a API exigir uma chave ou token armazenado numa conexão de projeto do Foundry.

{
  "description": "REST API with connection-based auth",
  "tools": [
    {
      "type": "openapi",
      "openapi": {
        "name": "my-api",
        "spec": { "<paste OpenAPI spec object here>" },
        "auth": {
          "type": "connection",
          "security_scheme": {
            "project_connection_id": "<CONNECTION_NAME>"
          }
        }
      }
    }
  ]
}

Autenticação de identidade gerida:

Utilize este padrão quando a API de destino se autenticar através do Microsoft Entra ID. A identidade gerida do projeto Foundry chama a API em nome do agente. Certifique-se de que a identidade gerida tem o papel RBAC exigido no serviço alvo antes de usar este padrão.

{
  "description": "REST API with managed identity auth",
  "tools": [
    {
      "type": "openapi",
      "openapi": {
        "name": "my-api",
        "spec": { "<paste OpenAPI spec object here>" },
        "auth": {
          "type": "managed_identity",
          "security_scheme": {
            "audience": "<TARGET_SERVICE_AUDIENCE>"
          }
        }
      }
    }
  ]
}

from azure.ai.projects.models import OpenAPITool

tools = [
    OpenAPITool(
        name="my-api",
        spec={"<paste OpenAPI spec object here>"},
        auth={"type": "anonymous"},
    )
]

BinaryData specBytes = BinaryData.FromString("<OpenAPI spec JSON>");
ProjectsAgentTool tool = new OpenAPITool(
    new OpenApiFunctionDefinition(
        name: "my-api",
        spec: specBytes,
        openApiAuthentication: new OpenApiAnonymousAuthDetails()
    )
);

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "REST API via OpenAPI spec"
);

const tools = [
  {
    type: "openapi",
    openapi: {
      name: "my-api",
      spec: { /* paste OpenAPI spec object here */ },
      auth: {
        type: "anonymous",
      },
    },
  },
];

Autenticação baseada em teclas:

parameters:
  api_key:
    secret: true
    description: API key for the target service
resources:
  - kind: connection
    name: api-conn
    category: CustomKeys
    authType: CustomKeys
    target: https://api.example.com
    credentials:
      keys:
        key: "{{ api_key }}"
  - kind: toolbox
    name: openapi-tools
    description: OpenAPI key-auth toolbox
    tools:
      - type: openapi
        openapi:
          name: my-api
          spec:
            openapi: "3.0.1"
            info:
              title: "My API"
              version: "1.0"
            servers:
              - url: https://api.example.com/v1
            paths:
              /search:
                get:
                  operationId: search
                  parameters:
                    - name: query
                      in: query
                      required: true
                      schema:
                        type: string
                  responses:
                    "200":
                      description: OK
          auth:
            type: connection_auth
            connection_id: api-conn

{
  "description": "Delegate tasks to a specialist agent",
  "tools": [
    {
      "type": "a2a_preview",
      "name": "<AGENT_NAME>",
      "description": "<What this agent does>",
      "base_url": "<AGENT_BASE_URL>",
      "project_connection_id": "<CONNECTION_NAME>"
    }
  ]
}

from azure.ai.projects.models import A2APreviewTool

tools = [
    A2APreviewTool(
        name="<AGENT_NAME>",
        description="<What this agent does>",
        base_url="<AGENT_BASE_URL>",
        project_connection_id="<CONNECTION_NAME>",
    )
]

ProjectsAgentTool tool = new A2APreviewTool()
{
    ProjectConnectionId = "<CONNECTION_NAME>",
};

ToolboxVersion toolboxVersion = await toolboxClient.CreateToolboxVersionAsync(
    toolboxName: "my-toolbox",
    tools: [tool],
    description: "Delegate tasks to a specialist agent"
);

const tools = [
  {
    type: "a2a_preview",
    name: "<AGENT_NAME>",
    description: "<What this agent does>",
    base_url: "<AGENT_BASE_URL>",
    project_connection_id: "<CONNECTION_NAME>",
  },
];

resources:
  - kind: connection
    name: a2a-conn
    category: RemoteA2A
    authType: None
    target: https://your-remote-agent.azurecontainerapps.io
  - kind: toolbox
    name: a2a-tools
    description: Agent-to-Agent toolbox
    tools:
      - type: a2a_preview
        project_connection_id: a2a-conn