Hosted mcp tool

hosted-mcp-tool

Use HostedMCPTool when you want OpenAI's infrastructure to handle MCP tool calls to a publicly reachable server, without round-trips to your Python process. Works only with OpenAI Responses API models.

When to apply

When the MCP server is publicly accessible and you want to avoid the latency of local tool execution. The model lists and calls tools server-side.

Core snippet

import asyncio
from agents import Agent, HostedMCPTool, Runner

async def main():
    agent = Agent(
        name="MCP Assistant",
        instructions="Use the available tools to answer questions.",
        tools=[
            HostedMCPTool(
                tool_config={
                    "type": "mcp",
                    "server_label": "my_server",
                    "server_url": "https://my-mcp-server.example.com/mcp",
                    "require_approval": "never",
                }
            )
        ],
    )
    result = await Runner.run(agent, "What tools are available?")
    print(result.final_output)

asyncio.run(main())

With approval required

HostedMCPTool(
    tool_config={
        "type": "mcp",
        "server_label": "dangerous_server",
        "server_url": "https://example.com/mcp",
        "require_approval": "always",  # Pause for human approval
    }
)

Key notes

• Server must be publicly reachable from OpenAI's infrastructure
• require_approval: "never" auto-approves; "always" triggers result.interruptions
• Does not support local/private MCP servers (use MCPServerStdio or MCPServerSse instead)
• Only works with OpenAIResponsesModel-backed agents
• Tool listing and invocation happen server-side; no Python callback for individual tool calls