Wire an OpenAI mcp_approval_response to a Sume spend cap

Make the human approval and the spend cap the same number. When your app shows an mcp_approval_request for a paid Sume tool, show the dry_run estimate beside it, and have the approver confirm a dollar ceiling. Then use that ceiling as max_spend_usd on the real call so the model cannot spend beyond what the person saw.

OpenAI's MCP guide defines the round trip: the model output has an mcp_approval_request item, and you reply with an mcp_approval_response input item with approve and approval_request_id. Sume's optional spend controls are in tools and gates.

Sequence

The approval reply

The reply is an ordinary input item on the next request. This helper builds it from the request item. Chain it to the previous response so the model continues the same turn.

export function approvalReply(item: { id: string }, approve: boolean) {
  return {
    type: "mcp_approval_response",
    approval_request_id: item.id,
    approve,
  };
}

Gaps to know about

Approval in OpenAI's flow is about whether the call may proceed. It does not by itself set a cap on the Sume side, which is why the second step matters: Sume enforces max_spend_usd only when it is present. If the model controls the arguments, set the cap in your own proxy, or check the arguments in the approval request before you approve.