Gpt Image

---
name: gpt-image
description: Use GPT Image image tools through an Ace Data Cloud account in Claude Code, including setup and retrieval of the same task. Applies when the user requests GPT Image; it does not change Claude's chat model.
license: Apache-2.0
---

# GPT Image by Ace Data Cloud

This plugin supplies a service-specific workflow and connects the existing Ace MCP. Credentials come from the plugin's masked **Ace Data Cloud API Token** option, not from prompts or another vendor's environment variables.

## Setup and free check

If tools are absent or unauthorized, direct the user to `/plugin` → Installed → this plugin → Configure options. They need an Ace service API Token with the relevant allowed APIs and balance; see [the illustrated setup](../../README.md). Do not ask them to paste a token into this conversation or alter a global credential.

Call `gpt_image_list_models` with no arguments to read local model reference information. It does not submit a paid generation or prove balance/authorization for generation.

## Use the service

The related tool names are `gpt_image_list_models`, `gpt_image_generate`, `gpt_image_edit`, `gpt_image_get_task`, `gpt_image_list_tasks`. Inspect the connected tool's live input schema for currently supported model IDs and parameters; do not invent a model or rely on another service's defaults.

For generation/editing, collect the user's intended input, model and output limits, point out that this is paid through Ace Credits, and use the existing authorization for that exact request. If price or intended limits are unclear, resolve those before submitting. A setup request or a request to explain options does not authorize generation. Send only the prompt and any user-selected references needed for that operation, not the rest of the conversation or unrelated files.

When an operation returns only a task ID, retain that exact ID and poll the related get-task tool at the documented interval until success or failure. Do not submit again to check progress, automatically retry billable writes, or treat a task ID/partial preview as completion. If the result already contains terminal success and final media URLs, report those. Surface rate-limit, moderation, balance and internal errors accurately rather than claiming success.

The plugin starts the published `mcp-openai-pro==2026.11.6.1 --profile gpt-image` stdio profile. Its five-tool catalog excludes chat, audio, embeddings, DALL-E and Nano Banana. An explicit GPT Image model is required for creation/editing; no model fallback is added. The saved Token is passed only as ACEDATACLOUD_API_TOKEN to that process, never inserted into the conversation. Keep its key restricted to image/task APIs.

This plugin does not make GPT Image a native Claude chat model, and installation does not establish Anthropic directory approval. The full existing API reference remains maintained at https://github.com/AceDataCloud/Skills/tree/main/skills/gpt-image-2-5; read it only when an advanced action needs it.

## GPT Image 2.6 controls

The Backend API supports native GIF generation/editing for GPT Image 2.6 Flare/Sunburst (standard, reverse, official). This does not prove that the pinned MCP exposes every combination: inspect its live input schema and `gpt_image_list_models` reference before submitting. If the connected tool disallows the requested model/format, report that tool limitation; do not bypass validation or silently switch models.

For supported GIF requests, use `output_format=gif`. Start with `size=512x512`, `frame_count=9`, `frame_duration_ms=200`; omitted/auto size uses the preview preset and actual output dimensions may vary. Frames must be integers 2–16; duration must be an integer 20–655350 ms, rounded to the nearest 10ms when encoded. Frame controls are GIF-only; `n` counts GIF files, not frames. Do not combine GIF with masks, markup, custom compression, prompt rewriting or partial-image streaming. Size, frames and file count share a model pixel budget; static 4K limits do not establish GIF support.

For 2.6 generation/editing, `complexity=low/medium/high` controls planning independently of `quality`, defaults to high and has no separate surcharge; official billing still uses actual tokens. For static annotation edits, use `markup` only if the connected edit schema accepts it: a PNG with transparent pixels, matching the first reference after orientation, at most 16,777,216 pixels per image and 40MB for the annotation. Explain the marks in the prompt; do not combine with mask, GIF or prompt rewriting, and do not use `file_id` or `input_fidelity` for this workflow. Use only input forms declared by the tool.

Static PNG/JPEG/WebP generation/editing supports 4K concrete sizes such as `3840x2160`; neither `4K` nor `4096x4096` is a valid size. Follow the same-task polling procedure above.