Fixed-price MCP tools and max_spend_usd: $0.10 to $0.30 floors

Some of Sume's hosted MCP tools have one price per call, not a per-second price. For those, the server holds a fixed estimate in code and compares it with max_spend_usd on every call, before the request leaves. If your cap is below that estimate, the call fails with max_spend_exceeded. You do not need a dry run to learn it.

The fixed estimates

The code lists five paid tools with a fixed estimate: trending-videos_search and trending-research_search at $0.10 each, media-imports_create at $0.15, video-captions_create and video-caption-overlay_create at $0.20 for clips up to 60 seconds, and video-analyses_create at $0.30. The comments in the code label them as fixed public estimates.

So a cap of 0.1 passes a trending search but fails a captions call; a cap of 0.2 passes captions; 0.3 passes all five. The comparison is in micro-dollars, with the cap rounded down.

What a dry run returns

With dry_run set to true, these tools return an object of type mcp_paid_fixed_price_dry_run. It has would_submit set to false, the endpoint, an estimate with currency USD and the three amount fields, a label, and next steps. The cap check runs before the dry run branch, so a cap that is too low fails even in a preview. That is useful: you learn the cap is wrong at no cost.

Using it in an agent

Set max_spend_usd per tool from this table with a small margin, instead of one global number. A single low cap for all tools would block the more expensive ones, and a single high cap would stop protecting you. Because each call is checked alone, a loop that makes 50 search calls at $0.10 still spends $5, so count calls in your own loop too.

Always send idempotency_key. If the cap check fails, nothing was submitted, and the same key can be used after you raise the cap.