OpenAI MCP 服务器
除函数调用外,也可用远程MCP或安全MCP隧道扩展模型,连接并控制外部服务。调用可自动允许,也可要求开发者明确批准。
远程MCP是公共互联网中实现远程模型上下文协议的服务器;安全MCP隧道可连接本地或私有服务器,无需公开暴露。
本文针对Responses API。内置连接器仍兼容既有模型,弃用范围与示例见旧连接器。Agents API会话应采用MCP连接,支持托管服务端或沙箱发起连接。
安全MCP隧道
私有、本地部署或防火墙内服务器可用安全MCP隧道连接受支持的OpenAI产品,无需公开服务器。从openai/tunnel-client获取当前公开发行版。
快速开始
Responses请求使用 type: mcp。远程服务器设置 server_url,隧道连接本地服务器使用 tunnel_id。某些服务器还需在 authorization 提供OAuth访问令牌。
以下查询公共OpenAI文档MCP中的Responses流式文档。该服务只读、无需认证,因此示例跳过调用审批;涉及敏感数据时应保留审批。
来源示例1(bash)
程序中的自然语言译释29:
在OpenAI文档中搜索Responses API流式输出,并返回相关链接。
程序中的自然语言译释30:
搜索并读取公开的OpenAI文档。
## [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-6-astra",
"tools": [
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "never"
}
],
"input": "Search the OpenAI docs for Responses API streaming and return the relevant links."
}'
来源示例2(javascript)
本段复用的中文释义:译释29、译释30。
// [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-6-astra",
tools: [
{
type: "mcp",
server_label: "openai_docs",
server_description: "Search and read the public OpenAI documentation.",
server_url: "https://developers.openai.com/mcp",
require_approval: "never",
},
],
input:
"Search the OpenAI docs for Responses API streaming and return the relevant links.",
});
console.log(resp.output_text);
来源示例3(python)
本段复用的中文释义:译释29、译释30。
## [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6-astra",
tools=[
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "never",
},
],
input="Search the OpenAI docs for Responses API streaming and return the relevant links.",
)
print(resp.output_text)
来源示例4(go)
本段复用的中文释义:译释29、译释30。
// [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("openai_docs")
tool.OfMcp.ServerDescription = openai.String("Search and read the public OpenAI documentation.")
tool.OfMcp.ServerURL = openai.String("https://developers.openai.com/mcp")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{OfMcpToolApprovalSetting: openai.String("never")}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-6-astra",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("Search the OpenAI docs for Responses API streaming and return the relevant links.")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例5(java)
本段复用的中文释义:译释29、译释30。
// [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-6-astra")
.input(
"Search the OpenAI docs for Responses API streaming and return the relevant links.")
.addTool(
Tool.Mcp.builder()
.serverLabel("openai_docs")
.serverDescription("Search and read the public OpenAI documentation.")
.serverUrl("https://developers.openai.com/mcp")
.requireApproval(Tool.Mcp.RequireApproval.McpToolApprovalSetting.NEVER)
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例6(csharp)
本段复用的中文释义:译释29。
// [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-6-astra" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "openai_docs",
serverUri: new Uri("https://developers.openai.com/mcp"),
toolCallApprovalPolicy: DefaultMcpToolCallApprovalPolicy.NeverRequireApproval
)
);
options.InputItems.Add(ResponseItem.CreateUserMessageItem("Search the OpenAI docs for Responses API streaming and return the relevant links."));
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例7(ruby)
本段复用的中文释义:译释29、译释30。
## [中文] 配置公开文档MCP的server_url并提交文档查询;此示例无敏感数据且跳过审批。
require "openai"
openai = OpenAI::Client.new
response = openai.responses.create(
model: "gpt-6-astra",
tools: [
{
type: "mcp",
server_label: "openai_docs",
server_description: "Search and read the public OpenAI documentation.",
server_url: "https://developers.openai.com/mcp",
require_approval: "never"
}
],
input: "Search the OpenAI docs for Responses API streaming and return the relevant links."
)
puts(response.output_text)
必须信任连接的服务器,恶意服务器可能外传上下文中的敏感数据。使用前阅读风险与安全。
模型决定使用MCP时,先列出工具,在output产生 mcp_list_tools。以下示意只列搜索工具并缩短描述,实际还提供其他文档工具。
来源示例8(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcpl_68a6102a4968819c8177b05584dd627b0679e572a900e618",
"type": "mcp_list_tools",
"server_label": "openai_docs",
"tools": [
{
"annotations": {
"readOnlyHint": true,
"destructiveHint": false
},
"description": "Search the public OpenAI documentation.",
"input_schema": {
"$schema": "http://json-schema.org/draft-07/schema#",
"type": "object",
"properties": {
"query": {
"type": "string",
"minLength": 1
},
"limit": {
"type": "integer",
"minimum": 1,
"maximum": 50
},
"cursor": {
"type": "string"
}
},
"required": ["query"]
},
"name": "search_openai_docs"
}
]
}
随后调用工具会产生 mcp_call,记录发给工具的内容及工具返回结果。
来源示例9(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcp_68a6102d8948819c9b1490d36d5ffa4a0679e572a900e618",
"type": "mcp_call",
"approval_request_id": null,
"arguments": "{\"query\":\"Responses API streaming\",\"limit\":1}",
"error": null,
"name": "search_openai_docs",
"output": "{\"hits\":[{\"url\":\"https://developers.openai.com/api/docs/guides/streaming-responses\"}]}",
"server_label": "openai_docs"
}
工作原理
近期模型中多数支持Responses MCP,具体核对模型兼容性。OpenAI侧按导入工具定义和调用消耗的词元计费,不额外逐MCP调用收费。
步骤1:列出可用工具
在tools指定远程服务器后,API尝试获取工具清单。支持Streamable HTTP或HTTP/SSE传输。成功后产生mcp_list_tools,其tools列出成功导入的定义。
来源示例10(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcpl_68a6102a4968819c8177b05584dd627b0679e572a900e618",
"type": "mcp_list_tools",
"server_label": "openai_docs",
"tools": [
{
"annotations": {
"readOnlyHint": true,
"destructiveHint": false
},
"description": "Search the public OpenAI documentation.",
"input_schema": {
"$schema": "http://json-schema.org/draft-07/schema#",
"type": "object",
"properties": {
"query": {
"type": "string",
"minLength": 1
},
"limit": {
"type": "integer",
"minimum": 1,
"maximum": 50
},
"cursor": {
"type": "string"
}
},
"required": ["query"]
},
"name": "search_openai_docs"
}
]
}
只要该项保留在请求上下文中,后续轮次无需重新获取清单。建议在会话或工作流中保留,减少延迟。
筛选工具
服务器工具可能很多,增加成本和延迟。只需子集时用allowed_tools限制导入;这里仅导入search_openai_docs查文档链接。
来源示例11(bash)
本段复用的中文释义:译释29、译释30。
## [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-6-astra",
"tools": [
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "never",
"allowed_tools": ["search_openai_docs"]
}
],
"input": "Search the OpenAI docs for Responses API streaming and return the relevant links."
}'
来源示例12(javascript)
本段复用的中文释义:译释29、译释30。
// [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-6-astra",
tools: [
{
type: "mcp",
server_label: "openai_docs",
server_description: "Search and read the public OpenAI documentation.",
server_url: "https://developers.openai.com/mcp",
require_approval: "never",
allowed_tools: ["search_openai_docs"],
},
],
input:
"Search the OpenAI docs for Responses API streaming and return the relevant links.",
});
console.log(resp.output_text);
来源示例13(python)
本段复用的中文释义:译释29、译释30。
## [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6-astra",
tools=[
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "never",
"allowed_tools": ["search_openai_docs"],
}
],
input="Search the OpenAI docs for Responses API streaming and return the relevant links.",
)
print(resp.output_text)
来源示例14(go)
本段复用的中文释义:译释29、译释30。
// [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("openai_docs")
tool.OfMcp.ServerDescription = openai.String("Search and read the public OpenAI documentation.")
tool.OfMcp.ServerURL = openai.String("https://developers.openai.com/mcp")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{OfMcpToolApprovalSetting: openai.String("never")}
tool.OfMcp.AllowedTools = responses.ToolMcpAllowedToolsUnionParam{OfMcpAllowedTools: []string{"search_openai_docs"}}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-6-astra",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("Search the OpenAI docs for Responses API streaming and return the relevant links.")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例15(java)
本段复用的中文释义:译释29、译释30。
// [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
import java.util.List;
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-6-astra")
.input(
"Search the OpenAI docs for Responses API streaming and return the relevant links.")
.addTool(
Tool.Mcp.builder()
.serverLabel("openai_docs")
.serverDescription("Search and read the public OpenAI documentation.")
.serverUrl("https://developers.openai.com/mcp")
.requireApproval(Tool.Mcp.RequireApproval.McpToolApprovalSetting.NEVER)
.allowedToolsOfMcp(List.of("search_openai_docs"))
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例16(csharp)
本段复用的中文释义:译释29。
// [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-6-astra" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "openai_docs",
serverUri: new Uri("https://developers.openai.com/mcp"),
allowedTools: new McpToolFilter() { ToolNames = { "search_openai_docs" } },
toolCallApprovalPolicy: DefaultMcpToolCallApprovalPolicy.NeverRequireApproval
)
);
options.InputItems.Add(ResponseItem.CreateUserMessageItem("Search the OpenAI docs for Responses API streaming and return the relevant links."));
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例17(ruby)
本段复用的中文释义:译释29、译释30。
## [中文] 用allowed_tools限制只导入文档搜索工具,不开放服务器全部工具。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-6-astra",
input: "Search the OpenAI docs for Responses API streaming and return the relevant links.",
tools: [
{
type: :mcp,
server_label: "openai_docs",
server_description: "Search and read the public OpenAI documentation.",
server_url: "https://developers.openai.com/mcp",
require_approval: :never,
allowed_tools: ["search_openai_docs"]
}
]
)
puts(response.output_text)
步骤2:调用工具
模型根据上下文决定调用,API请求远程MCP并把输出放回上下文,产生mcp_call。
来源示例18(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcp_68a6102d8948819c9b1490d36d5ffa4a0679e572a900e618",
"type": "mcp_call",
"approval_request_id": null,
"arguments": "{\"query\":\"Responses API streaming\",\"limit\":1}",
"error": null,
"name": "search_openai_docs",
"output": "{\"hits\":[{\"url\":\"https://developers.openai.com/api/docs/guides/streaming-responses\"}]}",
"server_label": "openai_docs"
}
项中包含参数与output。模型可多次调用,因此一次API请求可能有多个项。失败时error含协议错误、工具执行错误或连接错误,参见MCP错误规范。
批准
默认在向连接器或远程服务器分享数据前请求批准,便于控制与观察共享内容,建议仔细审核并可记录。将快速开始的require_approval从never改为always,可体验审批。输出会出现mcp_approval_request:
来源示例19(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa",
"type": "mcp_approval_request",
"arguments": "{\"query\":\"Responses API streaming\",\"limit\":1}",
"name": "search_openai_docs",
"server_label": "openai_docs"
}
审核工具与参数,再创建新响应并加入mcp_approval_response。将示意响应ID和请求ID换为真实值;.NET示例从首个响应读取。每次批准仅对应一次调用,后续请求照样处理。
来源示例20(bash)
本段复用的中文释义:译释30。
## [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-6-astra",
"tools": [
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "always"
}
],
"previous_response_id": "resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa",
"input": [{
"type": "mcp_approval_response",
"approve": true,
"approval_request_id": "mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa"
}]
}'
来源示例21(javascript)
本段复用的中文释义:译释30。
// [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-6-astra",
tools: [
{
type: "mcp",
server_label: "openai_docs",
server_description: "Search and read the public OpenAI documentation.",
server_url: "https://developers.openai.com/mcp",
require_approval: "always",
},
],
previous_response_id: "resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa",
input: [
{
type: "mcp_approval_response",
approve: true,
approval_request_id:
"mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa",
},
],
});
console.log(resp.output_text);
来源示例22(python)
本段复用的中文释义:译释30。
## [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6-astra",
tools=[
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "always",
}
],
previous_response_id="resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa",
input=[
{
"type": "mcp_approval_response",
"approve": True,
"approval_request_id": "mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa",
}
],
)
print(resp.output_text)
来源示例23(go)
本段复用的中文释义:译释30。
// [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("openai_docs")
tool.OfMcp.ServerDescription = openai.String("Search and read the public OpenAI documentation.")
tool.OfMcp.ServerURL = openai.String("https://developers.openai.com/mcp")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{OfMcpToolApprovalSetting: openai.String("always")}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-6-astra",
PreviousResponseID: openai.String("resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa"),
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfInputItemList: responses.ResponseInputParam{
responses.ResponseInputItemParamOfMcpApprovalResponse("mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa", true),
}},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例24(java)
本段复用的中文释义:译释30。
// [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.ResponseInputItem;
import com.openai.models.responses.Tool;
import java.util.List;
String responseId = "resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa";
String approvalRequestId = "mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa";
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-6-astra")
.input(
ResponseCreateParams.Input.ofResponse(
List.of(
ResponseInputItem.ofMcpApprovalResponse(
ResponseInputItem.McpApprovalResponse.builder()
.approvalRequestId(approvalRequestId)
.approve(true)
.build()))))
.previousResponseId(responseId)
.addTool(
Tool.Mcp.builder()
.serverLabel("openai_docs")
.serverDescription("Search and read the public OpenAI documentation.")
.serverUrl("https://developers.openai.com/mcp")
.requireApproval(Tool.Mcp.RequireApproval.McpToolApprovalSetting.ALWAYS)
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例25(csharp)
本段复用的中文释义:译释29。
// [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-6-astra" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "openai_docs",
serverUri: new Uri("https://developers.openai.com/mcp"),
toolCallApprovalPolicy: DefaultMcpToolCallApprovalPolicy.AlwaysRequireApproval
)
);
// [中文] 步骤1:创建会请求工具批准的响应。
// 第 1 步:创建请求审批工具调用的响应。
options.InputItems.Add(ResponseItem.CreateUserMessageItem("Search the OpenAI docs for Responses API streaming and return the relevant links."));
ResponseResult response1 = await client.CreateResponseAsync(options);
McpToolCallApprovalRequestItem approvalRequest =
response1.OutputItems.OfType<McpToolCallApprovalRequestItem>().Single();
// [中文] 步骤2:批准该调用,再读取最终响应。
// 第 2 步:批准工具调用并获取最终响应。
options.PreviousResponseId = response1.Id;
options.InputItems.Clear();
options.InputItems.Add(
ResponseItem.CreateMcpApprovalResponseItem(approvalRequest.Id, approved: true)
);
ResponseResult response2 = await client.CreateResponseAsync(options);
Console.WriteLine(response2.GetOutputText());
来源示例26(ruby)
本段复用的中文释义:译释30。
## [中文] 批准指定的MCP请求;approval_request_id及previous_response_id必须来自真实响应。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-6-astra",
previous_response_id: "resp_682d498bdefc81918b4a6aa477bfafd904ad1e533afccbfa",
input: [
{
type: :mcp_approval_response,
approval_request_id: "mcpr_682d498e3bd4819196a0ce1664f8e77b04ad1e533afccbfa",
approve: true
}
],
tools: [
{
type: :mcp,
server_label: "openai_docs",
server_url: "https://developers.openai.com/mcp",
server_description: "Search and read the public OpenAI documentation.",
require_approval: :always
}
]
)
puts(response.output_text)
通过previous_response_id关联发起审批的响应,也可手动把前次输出作为下次输入,以精细控制上下文。
建立信任后,可在require_approval对象中列出可跳过审批的工具,或设never跳过整个服务器全部工具,以降低延迟。
来源示例27(bash)
程序中的自然语言译释70:
2025-03-26版MCP规范支持哪些传输协议?
## [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-6-astra",
"tools": [
{
"type": "mcp",
"server_label": "deepwiki",
"server_url": "https://mcp.deepwiki.com/mcp",
"require_approval": {
"never": {
"tool_names": ["ask_question", "read_wiki_structure"]
}
}
}
],
"input": "What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?"
}'
来源示例28(javascript)
本段复用的中文释义:译释70。
// [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-6-astra",
tools: [
{
type: "mcp",
server_label: "deepwiki",
server_url: "https://mcp.deepwiki.com/mcp",
require_approval: {
never: {
tool_names: ["ask_question", "read_wiki_structure"],
},
},
},
],
input:
"What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?",
});
console.log(resp.output_text);
来源示例29(python)
本段复用的中文释义:译释70。
## [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6-astra",
tools=[
{
"type": "mcp",
"server_label": "deepwiki",
"server_url": "https://mcp.deepwiki.com/mcp",
"require_approval": {
"never": {"tool_names": ["ask_question", "read_wiki_structure"]}
},
},
],
input="What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?",
)
print(resp.output_text)
来源示例30(go)
本段复用的中文释义:译释70。
// [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("deepwiki")
tool.OfMcp.ServerURL = openai.String("https://mcp.deepwiki.com/mcp")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{
OfMcpToolApprovalFilter: &responses.ToolMcpRequireApprovalMcpToolApprovalFilterParam{
Never: responses.ToolMcpRequireApprovalMcpToolApprovalFilterNeverParam{
ToolNames: []string{"ask_question", "read_wiki_structure"},
},
},
}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-6-astra",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例31(java)
程序中的自然语言译释130:
2025-03-26版MCP规范支持哪些传输协议?
// [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-6-astra")
.input("What transport protocols does the 2025-03-26 version of the MCP spec support?")
.addTool(
Tool.Mcp.builder()
.serverLabel("deepwiki")
.serverUrl("https://mcp.deepwiki.com/mcp")
.requireApproval(
Tool.Mcp.RequireApproval.McpToolApprovalFilter.builder()
.never(
Tool.Mcp.RequireApproval.McpToolApprovalFilter.Never.builder()
.addToolName("ask_question")
.addToolName("read_wiki_structure")
.build())
.build())
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例32(csharp)
本段复用的中文释义:译释70。
// [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-6-astra" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "deepwiki",
serverUri: new Uri("https://mcp.deepwiki.com/mcp"),
toolCallApprovalPolicy: new CustomMcpToolCallApprovalPolicy
{
ToolsNeverRequiringApproval = new McpToolFilter
{
ToolNames = { "ask_question", "read_wiki_structure" },
},
}
)
);
options.InputItems.Add(
ResponseItem.CreateUserMessageItem(
"What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?"
)
);
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例33(ruby)
本段复用的中文释义:译释130。
## [中文] 配置部分工具免审批;这里是策略配置,不等于所有数据都适合外传。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-6-astra",
input: "What transport protocols does the 2025-03-26 version of the MCP spec support?",
tools: [
{
type: :mcp,
server_label: "deepwiki",
server_url: "https://mcp.deepwiki.com/mcp",
require_approval: {
never: { tool_names: ["ask_question", "read_wiki_structure"] }
}
}
]
)
puts(response.output_text)
身份验证
公共文档MCP无需认证,其他服务通常使用OAuth访问令牌,通过authorization传入。Stripe示例:
来源示例34(bash)
程序中的自然语言译释71:
创建金额为20美元的支付链接。
## [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-6-astra",
"input": "Create a payment link for $20",
"tools": [
{
"type": "mcp",
"server_label": "stripe",
"server_url": "https://mcp.stripe.com",
"authorization": "$STRIPE_OAUTH_ACCESS_TOKEN"
}
]
}'
来源示例35(javascript)
本段复用的中文释义:译释71。
// [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-6-astra",
input: "Create a payment link for $20",
tools: [
{
type: "mcp",
server_label: "stripe",
server_url: "https://mcp.stripe.com",
authorization: "$STRIPE_OAUTH_ACCESS_TOKEN",
},
],
});
console.log(resp.output_text);
来源示例36(python)
本段复用的中文释义:译释71。
## [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
import os
from openai import OpenAI
client = OpenAI()
authorization = os.environ["STRIPE_OAUTH_ACCESS_TOKEN"]
resp = client.responses.create(
model="gpt-6-astra",
input="Create a payment link for $20",
tools=[
{
"type": "mcp",
"server_label": "stripe",
"server_url": "https://mcp.stripe.com",
"authorization": authorization,
}
],
)
print(resp.output_text)
来源示例37(go)
本段复用的中文释义:译释71。
// [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
package main
import (
"context"
"fmt"
"os"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
authorization := os.Getenv("STRIPE_OAUTH_ACCESS_TOKEN")
if authorization == "" {
panic("STRIPE_OAUTH_ACCESS_TOKEN is required")
}
client := openai.NewClient()
tool := responses.ToolParamOfMcp("stripe")
tool.OfMcp.ServerURL = openai.String("https://mcp.stripe.com")
tool.OfMcp.Authorization = openai.String(authorization)
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-6-astra",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("Create a payment link for $20")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例38(java)
本段复用的中文释义:译释71。
// [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
String stripeAccessToken = System.getenv("STRIPE_OAUTH_ACCESS_TOKEN");
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-6-astra")
.input("Create a payment link for $20.")
.addTool(
Tool.Mcp.builder()
.serverLabel("stripe")
.serverUrl("https://mcp.stripe.com")
.authorization(stripeAccessToken)
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例39(csharp)
本段复用的中文释义:译释71。
// [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string authToken =
Environment.GetEnvironmentVariable("STRIPE_OAUTH_ACCESS_TOKEN")!;
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-6-astra" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "stripe",
serverUri: new Uri("https://mcp.stripe.com"),
authorizationToken: authToken
)
);
options.InputItems.Add(
ResponseItem.CreateUserMessageItem("Create a payment link for $20")
);
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例40(ruby)
本段复用的中文释义:译释71。
## [中文] 通过authorization给Stripe MCP提供令牌,并请求创建支付链接;真实运行具有外部副作用。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-6-astra",
input: "Create a payment link for $20.",
tools: [
{
type: :mcp,
server_label: "stripe",
server_url: "https://mcp.stripe.com",
authorization: ENV.fetch("STRIPE_OAUTH_ACCESS_TOKEN")
}
]
)
puts(response.output_text)
为避免敏感令牌泄漏,Responses不保存authorization值,返回的Response对象也不显示。因此每次创建响应都必须重新提供。
旧连接器
对于2026年9月1日之后发布的模型,connector_id 已弃用,应以server_url连接远程MCP,或tunnel_id经安全隧道连接本地MCP。已有模型保留连接器支持,以下使用早于截止日的gpt-5.2。
Responses内置有限的第三方连接器,可引入Dropbox、Gmail等上下文。与远程MCP作用相近,但以唯一connector_id代替server_url。应用必须通过authorization提供OAuth令牌。
来源示例41(bash)
程序中的自然语言译释72:
总结第二季度财报。
## [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-5.2",
"tools": [
{
"type": "mcp",
"server_label": "Dropbox",
"connector_id": "connector_dropbox",
"authorization": "<oauth access token>",
"require_approval": "never"
}
],
"input": "Summarize the Q2 earnings report."
}'
来源示例42(javascript)
本段复用的中文释义:译释72。
// [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-5.2",
tools: [
{
type: "mcp",
server_label: "Dropbox",
connector_id: "connector_dropbox",
authorization: "<oauth access token>",
require_approval: "never",
},
],
input: "Summarize the Q2 earnings report.",
});
console.log(resp.output_text);
来源示例43(python)
本段复用的中文释义:译释72。
## [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
import os
from openai import OpenAI
client = OpenAI()
connector_authorization = os.environ["OPENAI_CONNECTOR_AUTHORIZATION"]
resp = client.responses.create(
model="gpt-5.2",
tools=[
{
"type": "mcp",
"server_label": "Dropbox",
"connector_id": "connector_dropbox",
"authorization": connector_authorization,
"require_approval": "never",
},
],
input="Summarize the Q2 earnings report.",
)
print(resp.output_text)
来源示例44(go)
本段复用的中文释义:译释72。
// [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("Dropbox")
tool.OfMcp.ConnectorID = "connector_dropbox"
tool.OfMcp.Authorization = openai.String("<oauth access token>")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{OfMcpToolApprovalSetting: openai.String("never")}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-5.2",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("Summarize the Q2 earnings report.")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例45(java)
本段复用的中文释义:译释72。
// [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
String oauthAccessToken = "<oauth access token>";
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-5.2")
.input("Summarize the Q2 earnings report.")
.addTool(
Tool.Mcp.builder()
.serverLabel("Dropbox")
.connectorId(Tool.Mcp.ConnectorId.of("connector_dropbox"))
.authorization(oauthAccessToken)
.requireApproval(Tool.Mcp.RequireApproval.McpToolApprovalSetting.NEVER)
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例46(csharp)
本段复用的中文释义:译释72。
// [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string dropboxToken =
Environment.GetEnvironmentVariable("DROPBOX_OAUTH_ACCESS_TOKEN")!;
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-5.2" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "Dropbox",
connectorId: McpToolConnectorId.Dropbox,
authorizationToken: dropboxToken,
toolCallApprovalPolicy: DefaultMcpToolCallApprovalPolicy.NeverRequireApproval
)
);
options.InputItems.Add(
ResponseItem.CreateUserMessageItem("Summarize the Q2 earnings report.")
);
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例47(ruby)
本段复用的中文释义:译释72。
## [中文] 使用旧connector_id和兼容的gpt-5.2,凭证由OAuth流程提供。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-5.2",
input: "Summarize the Q2 earnings report.",
tools: [
{
type: :mcp,
server_label: "Dropbox",
connector_id: "connector_dropbox",
authorization: "<oauth access token>",
require_approval: :never
}
]
)
puts(response.output_text)
可用连接器
| 应用 | ID |
|---|---|
| Dropbox | connector_dropbox |
| Gmail | connector_gmail |
| Google Calendar | connector_googlecalendar |
| Google Drive | connector_googledrive |
| Microsoft Teams | connector_microsoftteams |
| Outlook Calendar | connector_outlookcalendar |
| Outlook Email | connector_outlookemail |
| SharePoint | connector_sharepoint |
来源优先提供尚无官方远程MCP的服务。GitHub已有官方MCP,可向server_url传 https://api.githubcopilot.com/mcp/。
为连接器授权
OAuth客户端注册和授权由应用单独完成,再把访问令牌放入authorization。测试可用Google OAuth 2.0 Playground生成临时令牌,在Step 1: Select and authorize APIs(选择并授权API)输入以下作用域以读取日历事件:
来源示例48(text):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
https://www.googleapis.com/auth/calendar.events
以Google账户授权后,在Step 2: Exchange authorization code for tokens(用授权码交换令牌)取得令牌,用日历连接器测试:
来源示例49(bash)
## [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
curl https://api.openai.com/v1/responses \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "gpt-5.2",
"tools": [
{
"type": "mcp",
"server_label": "google_calendar",
"connector_id": "connector_googlecalendar",
"authorization": "ya29.A0AS3H6...",
"require_approval": "never"
}
],
"input": "What is on my Google Calendar for today?"
}'
来源示例50(javascript)
程序中的自然语言译释73:
我今天的Google日历里有什么安排?
// [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.responses.create({
model: "gpt-5.2",
tools: [
{
type: "mcp",
server_label: "google_calendar",
connector_id: "connector_googlecalendar",
authorization: "ya29.A0AS3H6...",
require_approval: "never",
},
],
input: "What's on my Google Calendar for today?",
});
console.log(resp.output_text);
来源示例51(python)
本段复用的中文释义:译释73。
## [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
import os
from openai import OpenAI
client = OpenAI()
authorization = os.environ["GOOGLE_CALENDAR_OAUTH_ACCESS_TOKEN"]
resp = client.responses.create(
model="gpt-5.2",
tools=[
{
"type": "mcp",
"server_label": "google_calendar",
"connector_id": "connector_googlecalendar",
"authorization": authorization,
"require_approval": "never",
},
],
input="What's on my Google Calendar for today?",
)
print(resp.output_text)
来源示例52(go)
本段复用的中文释义:译释73。
// [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/responses"
)
func main() {
client := openai.NewClient()
tool := responses.ToolParamOfMcp("google_calendar")
tool.OfMcp.ConnectorID = "connector_googlecalendar"
tool.OfMcp.Authorization = openai.String("<oauth access token>")
tool.OfMcp.RequireApproval = responses.ToolMcpRequireApprovalUnionParam{OfMcpToolApprovalSetting: openai.String("never")}
response, err := client.Responses.New(context.Background(), responses.ResponseNewParams{
Model: "gpt-5.2",
Tools: []responses.ToolUnionParam{tool},
Input: responses.ResponseNewParamsInputUnion{OfString: openai.String("What's on my Google Calendar for today?")},
})
if err != nil {
panic(err)
}
fmt.Println(response.OutputText())
}
来源示例53(java)
本段复用的中文释义:译释73。
// [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.responses.ResponseCreateParams;
import com.openai.models.responses.Tool;
String oauthAccessToken = "<oauth access token>";
ResponseCreateParams params =
ResponseCreateParams.builder()
.model("gpt-5.2")
.input("What's on my Google Calendar for today?")
.addTool(
Tool.Mcp.builder()
.serverLabel("google_calendar")
.connectorId(Tool.Mcp.ConnectorId.of("connector_googlecalendar"))
.authorization(oauthAccessToken)
.requireApproval(Tool.Mcp.RequireApproval.McpToolApprovalSetting.NEVER)
.build())
.build();
client.responses().create(params).output().stream()
.flatMap(item -> item.message().stream())
.flatMap(message -> message.content().stream())
.flatMap(content -> content.outputText().stream())
.forEach(text -> System.out.println(text.text()));
来源示例54(csharp)
本段复用的中文释义:译释73。
// [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
using OpenAI.Responses;
#pragma warning disable OPENAI001
string authToken =
Environment.GetEnvironmentVariable("GOOGLE_CALENDAR_OAUTH_ACCESS_TOKEN")!;
string key = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
ResponsesClient client = new(key);
CreateResponseOptions options = new() { Model = "gpt-5.2" };
options.Tools.Add(
ResponseTool.CreateMcpTool(
serverLabel: "google_calendar",
connectorId: McpToolConnectorId.GoogleCalendar,
authorizationToken: authToken,
toolCallApprovalPolicy: DefaultMcpToolCallApprovalPolicy.NeverRequireApproval
)
);
options.InputItems.Add(
ResponseItem.CreateUserMessageItem("What's on my Google Calendar for today?")
);
ResponseResult response = await client.CreateResponseAsync(options);
Console.WriteLine(response.GetOutputText());
来源示例55(ruby)
本段复用的中文释义:译释73。
## [中文] 用OAuth令牌连接Google Calendar,提交查看今日日程的请求。
require "openai"
client = OpenAI::Client.new
response = client.responses.create(
model: "gpt-5.2",
input: "What's on my Google Calendar for today?",
tools: [
{
type: :mcp,
server_label: "google_calendar",
connector_id: "connector_googlecalendar",
authorization: "<oauth access token>",
require_approval: :never
}
]
)
puts(response.output_text)
连接器调用同样返回mcp_call,本例参数和响应均为JSON字符串。
来源示例56(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
{
"id": "mcp_68a62ae1c93c81a2b98c29340aa3ed8800e9b63986850588",
"type": "mcp_call",
"approval_request_id": null,
"arguments": "{\"time_min\":\"2025-08-20T00:00:00\",\"time_max\":\"2025-08-21T00:00:00\",\"timezone_str\":null,\"max_results\":50,\"query\":null,\"calendar_id\":null,\"next_page_token\":null}",
"error": null,
"name": "search_events",
"output": "{\"events\": [{\"id\": \"2n8ni54ani58pc3ii6soelupcs_20250820\", \"summary\": \"Home\", \"location\": null, \"start\": \"2025-08-20T00:00:00\", \"end\": \"2025-08-21T00:00:00\", \"url\": \"https://www.google.com/calendar/event?eid=Mm44bmk1NGFuaTU4cGMzaWk2c29lbHVwY3NfMjAyNTA4MjAga3doaW5uZXJ5QG9wZW5haS5jb20&ctz=America/Los_Angeles\", \"description\": \"\\n\\n\", \"transparency\": \"transparent\", \"display_url\": \"https://www.google.com/calendar/event?eid=Mm44bmk1NGFuaTU4cGMzaWk2c29lbHVwY3NfMjAyNTA4MjAga3doaW5uZXJ5QG9wZW5haS5jb20&ctz=America/Los_Angeles\", \"display_title\": \"Home\"}], \"next_page_token\": null}",
"server_label": "Google_Calendar"
}
各连接器可用工具
工具取决于OAuth令牌的作用域,以下保留来源的工具及权限对应。
Dropbox
| 工具 | 说明 | Scopes |
|---|---|---|
| search | 搜索匹配查询的文件 | files.metadata.read, account_info.read |
| fetch | 按路径取文件,可选原始下载 | files.content.read |
| search_files | 搜索并返回文件结果 | files.metadata.read, account_info.read |
| fetch_file | 读取文本或原始内容 | files.content.read, account_info.read |
| list_recent_files | 最近修改的可访问文件 | files.metadata.read, account_info.read |
| get_profile | 当前用户资料 | account_info.read |
Gmail
| 工具 | 说明 | Scopes |
|---|---|---|
| get_profile | 当前Gmail用户资料 | userinfo.email, userinfo.profile |
| search_emails | 按查询或标签搜索邮件 | gmail.modify |
| search_email_ids | 取匹配搜索的消息ID | gmail.modify |
| get_recent_emails | 最近收到的邮件 | gmail.modify |
| read_email | 读取邮件及正文 | gmail.modify |
| batch_read_email | 一次读取多封邮件 | gmail.modify |
Google Calendar
| 工具 | 说明 | Scopes |
|---|---|---|
| get_profile | 当前日历用户资料 | userinfo.email, userinfo.profile |
| search | 在可选时间窗搜索事件 | calendar.events |
| fetch | 单个事件详情 | calendar.events |
| search_events | 按条件查事件 | calendar.events |
| read_event | 按ID读事件 | calendar.events |
Google Drive
| 工具 | 说明 | Scopes |
|---|---|---|
| get_profile | 当前Drive用户资料 | userinfo.email, userinfo.profile |
| list_drives | 可访问的共享云端硬盘 | drive.readonly |
| search | 按查询搜索文件 | drive.readonly |
| recent_documents | 最近修改的文档 | drive.readonly |
| fetch | 下载文件内容 | drive.readonly |
Microsoft Teams
| 工具 | 说明 | Scopes |
|---|---|---|
| search | 搜索聊天和频道消息 | Chat.Read, ChannelMessage.Read.All |
| fetch | 按路径读取消息 | Chat.Read, ChannelMessage.Read.All |
| get_chat_members | 列出聊天成员 | Chat.Read |
| get_profile | 已认证用户资料 | User.Read |
Outlook Calendar
| 工具 | 说明 | Scopes |
|---|---|---|
| search_events | 按日期筛选事件 | Calendars.Read |
| fetch_event | 单个事件详情 | Calendars.Read |
| fetch_events_batch | 批量读取事件 | Calendars.Read |
| list_events | 列出日期范围内事件 | Calendars.Read |
| get_profile | 当前用户资料 | User.Read |
Outlook Email
| 工具 | 说明 | Scopes |
|---|---|---|
| get_profile | Outlook账户资料 | User.Read |
| list_messages | 从文件夹读取邮件 | Mail.Read |
| search_messages | 可选筛选条件搜索邮件 | Mail.Read |
| get_recent_emails | 最近收到的邮件 | Mail.Read |
| fetch_message | 按ID读取邮件 | Mail.Read |
| fetch_messages_batch | 批量读取邮件 | Mail.Read |
SharePoint
| 工具 | 说明 | Scopes |
|---|---|---|
| get_site | 按主机名及路径解析站点 | Sites.Read.All |
| search | 按关键字搜索SharePoint/OneDrive文档 | Sites.Read.All, Files.Read.All |
| list_recent_documents | 最近访问的文档 | Files.Read.All |
| fetch | 从Graph文件下载URL读取内容 | Files.Read.All |
| get_profile | 当前用户资料 | User.Read |
延迟加载MCP服务器中的工具
使用工具搜索时,在服务器定义设置defer_loading:true。模型仍看到服务器标签与描述以决定何时搜索,内部函数定义按需加载,可减少词元,对函数众多的服务器尤为有用。
来源示例57(json):以下保留协议载荷、配置、返回示意或语法定义;省略号及片段不代表完整可运行程序。
本段复用的中文释义:译释30。
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
// highlight-start:subtle
"defer_loading": true,
// highlight-end
"require_approval": "never"
}
风险与安全
连接器可能向OpenAI发送敏感数据,或让模型读取第三方敏感数据。远程MCP还有未经OpenAI验证的风险,可读写数据并执行动作,受各自第三方条款约束。发现恶意服务器可向security@openai.com报告。
提示注入
模型输入含用户提供内容、工具又可读敏感数据或操作服务时,应防范隐藏指令和工具滥用,使用适当保护。
敏感操作始终要求批准
通过require_approval及allowed_tools确保敏感操作进入审批。
工具调用和输出中的URL
请求输出提供的URL或嵌入其图像可能有危险,使用前确认域名与服务可信。
连接可信服务器
优先服务提供方自行托管的官方服务器,例如Stripe的mcp.stripe.com。第三方代理或聚合器需额外尽调,仔细核对数据使用。
记录和审查共享数据
服务器定义自己的工具,可能请求你不愿共享的信息,因此Responses默认每次调用审批。仔细审查共享范围,建立信任后再考虑跳过以降低延迟。建议记录发送数据并定期复查;来源说明store=true通常记录30天,组织启用ZDR时另有规则,也可在自己的系统保存日志。
恶意MCP可能在返回中藏指令,平台防护不能替代输入输出审查;只连接可信服务器。服务器工具行为也可能突然变化,产生意外或恶意行为。
ZDR与数据驻留的影响
MCP工具兼容ZDR及数据驻留,但第三方收到的数据受其自己的政策约束。例如欧洲驻留组织,OpenAI在向MCP发送前将客户内容推理及存储限制在欧洲;第三方是否也满足要求由应用负责核实。参见数据控制。
使用说明
来源表列出Responses、Chat Completions、Assistants入口;Build为1000 RPM,Launch and Grow为2000 RPM,并链接价格和数据要求。正文流程针对Responses,不能据此将其MCP请求结构直接用于其他接口。