Creating and editing share a single entry point, and the input image path is what switches between them, so there is no second tool to learn and no separate edit flow. The prompt optimizer is doing real work here — a short brief becomes a described scene before the image model sees it — and it is the first thing to turn off when the exact wording matters. What the server does not do is anything after the save: no listing, no deletion, no reopening an earlier result, so the file name you choose is the handle you keep.
A single-tool image server in front of Gemini, OpenAI or BytePlus Seedream, writing every result into an output directory and handing it back as a file resource.
- Text-to-image and image editing through one call: a prompt alone creates, a prompt plus an input image path edits the file at that path
- Prompt enhancement that fills in subject, lighting, camera and palette detail before the image model sees it — switched off with SKIP_PROMPT_ENHANCEMENT when the wording has to survive unchanged
- Three quality presets — fast, balanced and quality — each mapped to a different model route, set as a default or per call
- Framing decided before generation: aspect ratios from square to ultra-wide and ultra-tall, resolution up to 4K depending on provider and route, and the saved file's name
- Character consistency for follow-up images that keep the same subject, and blending for merging several inputs into one result
- Google Search grounding and world knowledge on the Gemini provider, for pictures that have to match something real
- The provider chosen per request as well as per server, so one configuration can serve all three
Node.js 22 or newer, and a key for whichever provider you use: GEMINI_API_KEY by default, OPENAI_API_KEY with IMAGE_PROVIDER=openai — which needs a verified OpenAI organization — or ARK_API_KEY for Seedream. IMAGE_OUTPUT_DIR must be an absolute path and is created if missing; without it, results land in ./output. The server starts without any key, and the error on the first call names the variable to set. MIT.
One command plus a key — npx -y mcp-image, then supply credentials