MCP server for image editing: restyle, cut out, upscale

An MCP server for image editing exposes edits as tools an agent can call on a picture that already exists. Sume's hosted MCP server at https://mcp.sume.com/mcp has three: generate_image with input_references edits or restyles reference images from a prompt, rmbg_create removes the background and returns a PNG with alpha, and image_upscale_create enlarges a still up to 4×. Each is a paid tool that takes public HTTPS image URLs and returns a job.

The facts come from Sume's MCP tools and gates and Image API pages, the Sume API reference for RMBG 1.0 and Image Upscale 1.0, and the tool contracts on Sume's current MCP server, read on 2026-09-27. The REST versions of the last two are covered in Remove background API and AI image upscaler API.

Which image edits can an agent run over Sume MCP?

Models whose input_references descriptor is {"min": 0, "max": 0} are text-to-image only and reject references; image-models_list shows each model's descriptors. Omit payload.model to route to sume/auto, or name a catalog model that takes references.

How does an agent call an image edit tool?

Sume's paid tools take { idempotency_key, payload }, with the request fields inside payload; tools_schema returns the live contract. Call once with dry_run: true for an admission and cost preview, then again without it to submit, and add max_spend_usd when you want a cap, enforced only when sent. Each submit answers with a job: the agent waits with jobs_wait, reads jobs_result, and never resubmits the paid create (MCP tool call timeouts on long-running video jobs covers the wait). This generate_image call restyles a photo:

{
  "idempotency_key": "watercolor-hero-001",
  "dry_run": true,
  "payload": {
    "prompt": "make this scene look like a watercolor painting",
    "aspect_ratio": "auto",
    "input_references": [
      {
        "type": "image_url",
        "image_url": { "url": "https://example.com/photo.jpg" }
      }
    ]
  }
}

Can an MCP image tool make a transparent background?

Not in one step. Sume's current generate_image contract says transparent or cutout output isn't available there: create or edit the still first, then run rmbg_create on its URL, which returns a PNG with alpha.

What does an image edit cost?

Background removal and upscaling have fixed rates on Sume's API pricing: $0.0225 per image to remove a background and $0.20 per image to upscale, each plus a 5.5% agent fee by default. A generate_image edit is priced per model: on Sume's current server, image-models_get returns a model's endpoint record with its pricing, and dry_run returns the estimate before anything is submitted. Image generation billing is all-or-nothing, so a failed or canceled generation is not billed.

What can't Sume's image tools do?

• Read files from your laptop. Hosted MCP can't, so every image goes in as a public HTTPS URL; localhost, private-network, and non-HTTPS references are rejected before submission.
• Run in a read-only session. With OAuth and Write off, all three tools return insufficient_scope; turn Write on at consent or connect with an API key.
• Generate a transparent still directly. The only transparent output is the rmbg_create cutout.
• Promise a look. How an edit comes out depends on the model, the prompt, and the references, so check the result before you use it. Sume MCP tools list lists the other tools an agent can chain.