Semantic Kernel MCPStreamableHttpPlugin with Sume: load_prompts False

To use Sume from Semantic Kernel, add Sume's hosted MCP endpoint (https://mcp.sume.com/mcp) with MCPStreamableHttpPlugin and set load_prompts to False. Microsoft's documentation, read on 2026-10-04, warns that prompt loading can hang when a server has none, and Sume's server exposes tools only. Then raise request_timeout above what your longest tool call needs, because one jobs_wait call on Sume can hold for 55 seconds.

What does Semantic Kernel offer?

The Semantic Kernel MCP plugins page lists three plugin types: MCPStdioPlugin, MCPSsePlugin and MCPStreamableHttpPlugin. You install the extra with pip install semantic-kernel[mcp]. The options are load_tools, which defaults to true, load_prompts, which also defaults to true, and request_timeout, for which the page does not state a unit.

Which settings suit Sume?

How should I authenticate?

For an unattended agent, use an API key, sent as Authorization: Bearer <SUME_API_KEY> or in an x-api-key header, as MCP OAuth and API keys describes. I have not verified how the plugin takes headers, so check Microsoft's page for the parameter before wiring it in. An API-key session sees the full tool set, so keep the number of functions the kernel can call small and require idempotency_key on every write and paid call.

OAuth suits an interactive app. A mcp:read grant hides write and paid tools, which is the right default for a copilot that only inspects jobs and usage.

What is the first call?

Start with mcp_health, which reports endpoint readiness, auth source and safety posture, then tools_list. For any generation, call the create tool and then jobs_wait and jobs_result, never a wait loop inside the create. The MCP quickstart walks through those calls, and Jobs and results covers statuses.

Plan the timeout for the slowest call you make, not the average. A wait slice that returns with jobs unfinished is normal, so call it again with the ids that remain rather than raising the timeout without limit.

How do I keep the agent safe?

Limit what the kernel may call. Sume's paid tools need an idempotency_key, and max_spend_usd caps a call only when it is sent, so have your own code, not the model, set the cap and build the key from the work. Start with read tools such as catalog_list and usage_get, then add one create tool once you have seen a full create, wait and result cycle succeed. That order keeps the first mistakes cheap.