OpenAI Agents SDK MCPServerManager with a Sume server

In the OpenAI Agents Python SDK, connect Sume as an MCPServerStreamableHttp server with a url and headers, and let MCPServerManager coordinate it with your other servers. The SDK page says the manager exposes the successfully connected subset, so confirm Sume is in that subset with tools_list.

SDK facts are from the OpenAI Agents SDK MCP page; Sume facts from MCP overview and OAuth and API keys, read 2026-09-30.

What server types does the SDK offer?

How do I connect Sume?

The Sume docs say an API key goes in Authorization: Bearer or x-api-key, and that API-key sessions see the full hosted tool set. The SDK page lists cache_tools_list among the constructor options.

import asyncio
import os

from agents.mcp import MCPServerStreamableHttp


async def main():
    async with MCPServerStreamableHttp(
        name="sume",
        params={
            "url": "https://mcp.sume.com/mcp",
            "headers": {"x-api-key": os.environ["SUME_API_KEY"]},
        },
        cache_tools_list=True,
    ) as sume:
        tools = await sume.list_tools()
        print([t.name for t in tools])


asyncio.run(main())

How do I know Sume actually connected?

Because a failed server can simply drop out of the connected subset, have the agent call mcp_health, which the docs say confirms endpoint, auth source and safety posture, and then tools_list. If a paid tool is missing on an OAuth session, Write was left off at consent.

Which tool ids should stay off the agent?

Paid tools such as generate_video need an idempotency_key, and the docs prefer dry_run or generation_admission_preview before an expensive burst. Give a research agent OAuth read-only access instead of a key.

Does a read-only session change the tool list?

Sume's docs say tools_list shows only the tools visible to the session, and OAuth mcp:read hides mutating and paid tools. A research-only agent on a read-only session therefore sees fewer Sume tools and cannot spend.