MCP generate_image: model defaults to sume/auto, spend preview

generate_image is the Sume MCP tool for images. Its payload matches the REST schema for POST /v1/images, payload.model is optional and defaults to sume/auto, and, as a paid-generation tool, it goes through an admission preview before the job is submitted. The tool list shows generate_image, image-models_list and image-models_get.

The three image tools

Source: the MCP section of the repo's docs/api/images.md.

What an agent should do

Have the agent call image-models_list before it pins an id, so it does not guess a model that is not in the catalog. A drift test keeps the MCP payload keys in step with the REST request schema, so a field that works on REST works on MCP.

The older name image-generations_create still resolves when called, but the tool list does not show it.

{
  "payload": {
    "model": "google/nano-banana-2.1",
    "prompt": "hero image for a bakery site, warm light",
    "aspect_ratio": "16:9"
  }
}

Cost control

Because the tool is gated, the preview shows the estimate before spend. Give the agent a pinned, cheaper id for drafts and keep the expensive rows for finals. Current credits are on the rate card.