MCP Service¶
OpenAI、Anthropic 或自建 Agent 都可以把 dxgate 暴露的 MCP endpoint 当成远程工具服务器。dxgate 负责路由、tool federation、认证、RBAC、限流和审计；模型负责决定何时调用工具。
flowchart TB
app["Application"] --> openai["OpenAI"]
app --> anthropic["Anthropic"]
openai --> gateway["dxgate"]
anthropic --> gateway
gateway -->|"/mcp"| search["search MCP"]
gateway -->|"/mcp"| calendar["calendar MCP"]
1. MCP Server¶
MCP Server 仍用普通 Kubernetes Service 暴露。协议类型由引用它的 DxgateService.spec.mcp 声明，因此不需要旧 DxgateBackend：
apiVersion: v1
kind: Service
metadata:
name: search-mcp
spec:
selector:
app: search-mcp
ports:
- port: 8080
targetPort: 8080
---
apiVersion: v1
kind: Service
metadata:
name: calendar-mcp
spec:
selector:
app: calendar-mcp
ports:
- port: 8080
targetPort: 8080
2. MCP federation¶
一个 DxgateService.spec.mcp 可以把多个 Service 联合成一个入口：
apiVersion: networking.dubbo.apache.org/v1alpha3
kind: DxgateService
metadata:
name: tools
spec:
mcp:
targets:
- name: search
static:
backendRef:
name: search-mcp
port: 8080
tools: [search]
- name: calendar
static:
backendRef:
name: calendar-mcp
port: 8080
tools: [calendar]
每个 target 的 backendRef 是普通 Kubernetes Service。
3. HTTPRoute¶
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
name: mcp
spec:
parentRefs:
- name: dxgate-proxy
namespace: dubbo-system
rules:
- matches:
- path:
type: PathPrefix
value: /mcp
backendRefs:
- name: tools
group: networking.dubbo.apache.org
kind: DxgateService
远程入口为 https://ai.example.com/mcp。
运行行为¶
/mcp与/mcp/*按 MCP 处理。initialize、tools/list等未指定工具的请求可访问全部 target。tools/call按tools声明选择 target。mcp-session-id把后续请求固定到首次选中的 target。tools/list、prompts/list、resources/list与模板列表会跨 target 聚合并处理分页。- 重名工具暴露为
{target}__{name}，调用时自动还原并发往对应 target。
4. 直接验证¶
列出 federation 中的工具：
curl https://ai.example.com/mcp \
-H 'Content-Type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'
调用工具：
curl https://ai.example.com/mcp \
-H 'Content-Type: application/json' \
-d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search","arguments":{"query":"order 12345"}}}'
5. OpenAI 使用远程 MCP¶
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-5",
input="查询订单 12345 的状态",
tools=[{
"type": "mcp",
"server_label": "company-tools",
"server_url": "https://ai.example.com/mcp",
}],
)
print(response.output_text)
6. Anthropic 使用同一个 MCP¶
import anthropic
client = anthropic.Anthropic()
response = client.beta.messages.create(
model="claude-opus-4-6",
max_tokens=1024,
messages=[{"role": "user", "content": "查询订单 12345 的状态"}],
mcp_servers=[{
"type": "url",
"name": "company-tools",
"url": "https://ai.example.com/mcp",
}],
tools=[{
"type": "mcp_toolset",
"mcp_server_name": "company-tools",
}],
betas=["mcp-client-2025-11-20"],
)
Claude MCP connector 当前需要 mcp-client-2025-11-20 beta。完整策略见统一 DxgateService API。生产环境应在公网 MCP endpoint 上启用 TLS 和客户端认证。