Microsoft Agent Framework hosted MCP tool for Sume (Python)

Short answer

Create the tool with client.get_mcp_tool(name=..., url="https://mcp.sume.com/mcp", headers=..., approval_mode=..., allowed_tools=[...]) and pass it in tools=[...] to Agent. The Microsoft Learn page, last updated 2026-09-23, documents that call shape for Python, including headers, approval_mode values never_require and always_require, and allowed_tools.

The detail that matters for Sume: a hosted MCP tool is executed by the backing service, not by your process. The page says the MCP server is hosted and managed by Foundry for the Foundry client, and for the OpenAI client says the tool is executed remotely by OpenAI. So the Sume API key you put in headers travels to that service.

What the Microsoft page documents

Approval and filtering are the controls you have. Both are worth setting deliberately when the server behind the tool can spend money.

The agent

This is the page's own pattern with Sume substituted for the example server. It allows only discovery and job-reading tools, so a first run cannot start paid generation; because the allowlist is read-only, the sample sets never_require so the run completes without an approval loop. Set FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_MODEL as the page describes, sign in with the Azure CLI, and export SUME_API_KEY.

import asyncio, os
from agent_framework import Agent
from agent_framework.foundry import FoundryChatClient
from azure.identity.aio import AzureCliCredential

async def main():
    async with AzureCliCredential() as credential:
        client = FoundryChatClient(credential=credential)
        sume = client.get_mcp_tool(
            name="Sume",
            url="https://mcp.sume.com/mcp",
            headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
            approval_mode="never_require",
            allowed_tools=["mcp_health", "tools_list", "jobs_status", "jobs_wait"],
        )
        async with Agent(
            client=client,
            name="SumeAgent",
            instructions="Use the Sume tools to report endpoint health and job status.",
            tools=[sume],
        ) as agent:
            result = await agent.run("Call mcp_health, then tools_list, and summarize.")
            print(result.text)

asyncio.run(main())

Opening the allowlist later

To let the agent submit work, add the generate tool you intend to use to allowed_tools, switch approval_mode to always_require (and handle the approval requests your host surfaces), and remember that Sume's hosted MCP requires an idempotency_key on every mutating or paid call. Ask for dry_run=true first, and pass max_spend_usd so a looping agent has a ceiling.

For a render that outlasts one wait, jobs_wait holds at most 55 seconds per call and answers wait_slice_expired when the slice ends; the agent should call it again with the same ids instead of resubmitting.

What Sume does and does not do

Sume's hosted MCP serves read tools to an OAuth session by default and the full tool set to an API-key session. It does not accept an OAuth token as an API key, and it has no paid OAuth scope.

Sume cannot see or limit what the agent host does with the key beyond the tool gates: use a key created for this integration so it can be rotated alone.