GPT Image MCP server: use GPT Image 2.5 in Claude or Cursor

To use GPT Image from Claude or Cursor, connect a remote MCP server that runs OpenAI's GPT Image models: Higgsfield, Runway and Sume each list GPT Image 2 for MCP clients, and Codex has GPT Image built in. On Sume's hosted server, generate_image with openai/gpt-image-2.5 runs ChatGPT Image 2.5, with up to 16 reference images and an optional mask, and openai/gpt-image-2 runs ChatGPT Image 2.

Sume's side comes from its MCP overview, MCP quickstart and Image API docs, the API reference and the hosted tool's code; the other options come from their own pages. All were read on 2026-09-28. The basics page says hosted MCP still works but is not the primary path today; from a backend, GPT Image 2.5 API on Sume covers POST /v1/images.

Which MCP servers run GPT Image?

Three remote servers read for this post name GPT Image, and Codex has it built in. Each bills its own account:

How do I add a GPT Image MCP server to Claude Code?

For Sume, run claude mcp add --transport http sume https://mcp.sume.com/mcp, then claude mcp login sume. The default sign-in is read-only (mcp:read), and generate_image returns insufficient_scope until you turn Write on at consent. An API-key session sees the full hosted tool set.

Then name the model inside payload; without payload.model, generate_image routes to sume/auto. Paid tools need an idempotency_key, and dry_run: true previews admission and cost without submitting:

{
  "idempotency_key": "gpt-image-label-001",
  "dry_run": true,
  "payload": {
    "model": "openai/gpt-image-2.5",
    "prompt": "put the label text 'Morning Blend' on this coffee bag",
    "quality": "high",
    "aspect_ratio": "auto",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/bag.png" } }
    ]
  }
}

What can GPT Image 2.5 do over Sume's MCP?

The model options are the ones GPT Image 2.5 API on Sume walks through: up to 16 references, an optional mask, and quality from low to max. The hosted tool adds its own rules, from its description and payload schema in current code:

• Read supported_parameters from image-models_list before pinning a setting; an unadvertised parameter returns a 400.
• Explicit pixel sizes are not served on this tool: pick an aspect_ratio, and upscale with image_upscale_create if you need more pixels.
• The payload takes mask_url for GPT Image 2.5 edits.
• With references, send aspect_ratio: "auto" to match the reference; omitting the field is not the same.
• Reference and mask URLs must be public HTTPS; hosted MCP cannot read files from your laptop.
• Results are Sume-hosted, signed URLs, so download the files you keep.

Can GPT Image make a transparent background over MCP?

Not in one step on Sume's server. The hosted tool's own description lists background as not served yet and says transparent output is not available on that surface. The route it gives is to generate the still, then cut it out with rmbg_create, which in current code returns a PNG with alpha. Background removal is listed at $0.0225 per image on API pricing; AI image generator with a transparent background covers prompting for a clean cutout.

How much does GPT Image cost over MCP?

On Sume, image billing is all-or-nothing: a completed generation is billed in full, and a failed or canceled one is not billed. The endpoint pricing lines from GET /v1/images/models are the amount charged, Sume's margin included. GPT Image 2.5 is priced by tokens, so size and quality move the amount, and auto quality reserves max; run dry_run first and send max_spend_usd to cap a call. In Codex, built-in image generation counts toward your general Codex usage limits instead.