Best AI Agent Skills for Image Generation (2026)

Every "generate images from Claude Code" guide skips the question that decides which option you should pick: where does the pixel get made, and who pays for it?

Three answers, and they lead to completely different setups. A local GPU running ComfyUI, where generation is free after the hardware and you own the models. A hosted API from Google, OpenAI, MiniMax or Replicate, where setup takes two minutes and every image bills to a key. Or a desktop application like Photoshop that the agent drives through a scripting bridge, where the image already exists and you are automating the edit.

Aggregator listings put all of these in one list called "image generation skills" and let you find out about the API key when the first call fails. Here is the split, with what each option costs you in setup time versus running cost.

Quick comparison

Local generation: you own the whole pipeline

ComfyUI MCP Server

joenorton/comfyui-mcp-server is a lightweight Python MCP server that lets an agent drive a local ComfyUI instance. Apache-2.0, 398 stars, community tier and unreviewed.

The appeal is control. ComfyUI graphs let you pin an exact model, sampler, seed and post-processing chain, so the output is reproducible in a way no hosted API gives you. If you are generating a hundred product variants that must all match, this is the only category that reliably delivers that.

The cost is that the MCP server is the easy part. You need ComfyUI installed, working, and loaded with the checkpoints your workflow references. Last push was February 2026, so verify it against your current ComfyUI version before building anything on top.

Install: claude mcp add comfyui-mcp-server -- npx -y joenorton/comfyui-mcp-server · Deep dive: /skills/comfyui-mcp-server

ComfyUI Custom Node Skills

jtydhr88/comfyui-custom-node-skills goes one level deeper: a collection of Claude Code skills for writing ComfyUI custom nodes. This is for the case where the workflow you need does not exist yet and you are building the node.

267 stars, unreviewed, license listed as unknown. Narrow audience, but if you are in it, nothing else on this page substitutes.

Deep dive: /skills/comfyui-custom-node-skills

Hosted APIs: fast to start, metered forever

Nanobanana MCP Server

zhongweili/nanobanana-mcp-server wraps Google Gemini image generation with model selection handled for you. MIT, 390 stars, community tier.

This is the shortest path from nothing to an image appearing in your working directory. Get a Gemini API key, add the server, ask for the image. The trade is that you are renting: no local model, no offline use, and cost scales with volume rather than sitting flat.

Install: claude mcp add nanobanana-mcp-server -- npx -y zhongweili/nanobanana-mcp-server · Deep dive: /skills/nanobanana-mcp-server

Nano Banana Pro Prompts Recommend Skill

Worth being precise about what this is, because the name suggests otherwise. YouMind-OpenLab/nano-banana-pro-prompts-recommend-skill recommends prompts from a large curated library. It does not generate anything. Pair it with a server that does.

That pairing is more useful than it sounds. Prompt quality is the largest single variable in image output, and a vetted library beats improvising for anyone who is not a full-time prompt writer. 1,824 stars, license listed as unknown, pushed recently.

Deep dive: /skills/nano-banana-pro-prompts-recommend-skill

GPT Image 2 Skill

wuyoscar/gpt_image_2_skill bundles a prompt gallery, an image prompt library, an agentic skill and a CLI for OpenAI image generation. MIT, verified tier, 4,526 stars, pushed August 2026.

Note that this repo appears in our directory under two slugs from two capitalisations of the same project. They are the same codebase. We are cleaning that class of duplicate up; in the meantime, install either and ignore the other.

Deep dive: /skills/gpt-image-2-skill

Replicate AI Models

Replicate's own MCP server, Apache-2.0, 1,400 stars, verified. Its advantage over any single-vendor option is breadth: Flux, Whisper and a long catalogue of community models behind one token, billed by compute second.

Use this when you do not yet know which model you want. Swapping models is a string change rather than a new integration, which makes it the right choice for the exploration phase.

Install: claude mcp add replicate -- npx -y @replicate/mcp-server · Deep dive: /skills/replicate-models

MiniMax MCP

The official MiniMax server, MIT, 1,559 stars, verified. Covers image alongside other generative modes, which matters if your project needs stills and motion from the same vendor and billing account. Last push was May 2026, the oldest date among the hosted options here.

Deep dive: /skills/minimax-mcp

Driving software you already own

Photoshop Python API MCP Server

loonghao/photoshop-python-api-mcp-server connects an agent to Adobe Photoshop's scripting interface. MIT, 292 stars, pushed August 2026, community tier and unreviewed.

Different job from everything above. Nothing is generated. You are automating the boring half of a real design workflow: batch resizing, layer operations, export sets, applying the same treatment across forty files. For a designer who already lives in Photoshop, this saves more hours per week than any text-to-image server will.

Windows and macOS only in practice, since it talks to a locally installed Photoshop.

Deep dive: /skills/photoshop-python-api-mcp-server

MeiGen AI Design MCP

jau123/MeiGen-AI-Design-MCP aims at turning Claude Code into a local design assistant across a broader design surface than single-image generation. MIT, 1,699 stars, community tier and unreviewed.

Younger and less proven than the others here, and the documentation is thinner. Interesting if you want design-system-shaped output rather than one-off images.

Deep dive: /skills/meigen-ai-design-mcp

Picking one

Work backwards from the constraint that will bite you first.

• Reproducibility matters most (product shots, brand assets, anything that must match across a set): local ComfyUI. Nothing hosted gives you seed-level control you can rely on months later.
• You want an image in the next ten minutes: Nanobanana or GPT Image 2. Key, install, done.
• You do not know which model fits yet: Replicate, then narrow.
• The images already exist and the work is editing: the Photoshop bridge, and skip this whole category.
• Output quality is your bottleneck rather than plumbing: add the Nano Banana Pro prompt library on top of whichever generator you chose.

One thing that applies across all of them. Every hosted option here bills your key, and an agent that can generate images can generate them in a loop. Set a spend cap on the API key before the first run, not after the first surprise.

For the tool side of this question rather than the skill side, see our comparisons of the best AI image generators and image tools for designers, plus our full list for designers.

FAQ

Is there an official Anthropic skill for image generation?

No. Anthropic's own collection covers document formats, frontend design and skill authoring, but image generation is community territory. Everything on this page is third-party, which is why the trust column matters more here than on most of our cluster pages.

Skill or MCP server, which do I want?

Most of this category is MCP servers, because generating an image means calling out to a process or an API and a markdown instruction file cannot do that alone. The two entries that are skills in the strict sense, the Nano Banana prompt library and the ComfyUI node collection, both shape how you use a generator rather than being one.

Can I run any of this without a GPU or an API key?

Not for generation. Local generation needs the GPU; hosted generation needs the key. The Photoshop bridge is the only entry that does useful work with neither, since it operates on images you already have.

Which of these are security-reviewed?

Replicate, MiniMax and GPT Image 2 sit at verified tier. The rest are community and unreviewed. Since these servers hold API keys that cost money per call, read the source before installing, and give each one its own key so you can revoke it without breaking everything else.

Why do two GPT Image 2 entries appear in the directory?

Two capitalisations of the same GitHub repository were ingested as separate rows. Same project, same stars, same install target. It is a known duplicate class in our data and it is on the cleanup list.