MCP: read image-models_get, then dry-run generate_image first

Before a paid image call over hosted MCP, an agent should read the model with image-models_get, then call generate_image with dry_run: true and an idempotency_key, and submit only after the estimate looks right. The dry run previews admission and cost and does not submit a job.

The hosted MCP URL is https://mcp.sume.com/mcp. Under OAuth with only mcp:read, the write and paid tools are hidden, so a session needs mcp:write or an API key to see generate_image at all.

Step 1: read the model

image-models_list returns the catalog and image-models_get returns one model's endpoint row. Its single required argument is model_id, such as openai/gpt-image-2.5. Read the supported sizes, qualities and reference limits there instead of guessing, and call tools_schema with name: "generate_image" for the exact call contract. The Sume docs add that the catalog does not list sume/auto.

Step 2: dry run with a cap

Send the same arguments you plan to submit, with dry_run true. Name the model in payload.model when the user asked for a specific family, and omit it otherwise.

{
  "idempotency_key": "banner-draft-2026-10-05-001",
  "dry_run": true,
  "max_spend_usd": 1,
  "payload": {
    "model": "openai/gpt-image-2.5",
    "prompt": "Flat teal banner with a white paper plane",
    "image_size": "1920x640",
    "quality": "medium"
  }
}

Step 3: submit and wait

Read the preview. If it matches, call again with dry_run omitted or false and a fresh idempotency_key, then follow the job with jobs_wait and read the output. Retry a transport failure with the same key, so the same create is not paid twice. For a burst of calls, run generation_admission_preview once first to see balance and queue behavior.

What a dry run does not prove

A preview is an estimate at the moment you ask. Balance, queue and model availability can change before the real call, so keep max_spend_usd on the submit as well if you want a hard cap.

A rule for agent instructions

These four lines fit in a system prompt and cover the cases where an agent spends money by accident. The gates are in the docs, but an agent follows them only if its instructions say so. Test the instructions once with a dry run on a cheap model and read the preview aloud in the agent's own report, so a human can see the estimate before the real call.

• Call tools_schema for generate_image before the first create in a session.
• Always dry run first when the user has not confirmed a price.
• Keep one idempotency_key per intended image, and reuse it only on a retry of that same call.
• Report the Sume job id and the result URL, and never paste a token or signed URL into chat.