Best Image-Generation MCP Servers (2026, Ranked)

If you want to generate images from inside Claude, Cursor, or any MCP-compatible agent in 2026, you are no longer stuck with a web app and a tab-switch. A small but real ecosystem of image generation MCP servers exists now, and they differ significantly in what they can actually do.

The short answer: AgentBrush is the most feature-complete option for most builders, covering generation, editing, background removal, and style presets in one server. Scenario is the right call for game studios with trained style models. The open-source wrappers (MeiGen MCP, imagegen-mcp, GongRzhe's Flux bridge) are worth knowing about if you want to roll your own setup or experiment with raw model access. Below is the full breakdown.

What to look for in an image generation MCP server

The Model Context Protocol lets your agent call tools the same way it calls a file-read or a web search. An image MCP server exposes that capability, but "generate an image" is only the starting point. Here is what separates a thin API bridge from a tool that holds up in a real project:

• Style presets. Generating a pixel sprite and a product photo require completely different prompts and model settings. A preset handles that translation so your agent does not have to write a 200-word system prompt every time.
• Editing and masking. One-shot generation is useful; being able to repaint a region without regenerating the whole image is a workflow.
• Background removal. If you build apps, games, or marketing assets, you need transparent PNGs. Running the removal locally matters: no extra API call, no extra cost, no upload to a third-party.
• Reference consistency. Getting the same character, product, or face across multiple outputs without LoRA or fine-tuning.
• Model choice. Most servers are locked to one provider. Wider model support matters if you need a specific capability or price point.

Quick comparison

Retro Diffusion is a family of specialized pixel-art models available through Scenario's platform, not a separate standalone MCP server as of this writing. Check Scenario's model library for current availability.

1. AgentBrush: most complete for general use

AgentBrush runs on gpt-image-2 and is the only server in this list that combines generation, in-browser mask editing, local background removal, style presets, and reference-image consistency in one place. It works with Claude Code, Cursor, Codex CLI, Gemini CLI, and any MCP-compatible client.

The six presets (realistic, flat_illustration, pixel_art, isometric, logo, custom) are the practical backbone. When your agent calls agentbrush_generate, it picks the preset and hands back the image; no 200-word style prompt required. The custom preset takes a style description of your own, which works well when you want something specific like a ukiyo-e print or a brutalist poster.

Background removal deserves a separate mention. Most tools that offer this run it in the cloud, charge per image, and upload your asset to do it. AgentBrush runs background removal locally: it costs 0 tokens, uploads nothing, and outputs a PNG with a real alpha channel. If you are generating product cut-outs or game sprites at any volume, that is the difference between a line item and a non-event. The full workflow is in generating transparent-background PNGs.

# Generate a subject on a clean background
preset: flat_illustration · quality: low
"A friendly robot holding a wrench, flat coral background"

# Strip the background locally (0 tokens)
agentbrush_remove_background path/to/robot.png

The mask editor lives in the browser and lets you repaint specific regions without regenerating the whole image. You open it from your agent, draw the region, and the result comes back as a new file. Not every project needs this, but when you do need it, the alternative is regenerating until you get lucky.

Reference-image consistency is handled by passing existing images as inputs when calling agentbrush_generate. For characters or products you want to keep visually coherent across multiple outputs, this is the practical lever.

gpt-image-2 has no reliable native transparency. The two-step approach (generate on a clean background, then agentbrush_remove_background locally) is the dependable path to a real alpha-channel PNG, and the removal step costs 0 tokens.

Where AgentBrush is honest about its limits: it is single-model (gpt-image-2), so if you specifically need Midjourney's aesthetic or a ComfyUI workflow with custom nodes, this is not the right server. It is also commercial, so if your constraint is zero spend, the open-source options below are worth a look.

Plans run $6.99/mo for 100 tokens (Starter), $14.99/mo for 600 tokens (Pro), and $29.99/mo for 1,300 tokens (Power). Token costs are 1/5/20 at low/medium/high quality at 1024px square; portrait and landscape sizes cost 1.5x. Starter and Pro are hard-capped with no overage. Power is the only tier with overage at $0.04/token. More on the math in tokens, quality, and cost.

2. Scenario MCP: best for trained game styles

Scenario is a commercial platform built around custom-trained style models for games. You train a model on your own art, and the Scenario MCP server lets your agent call that trained model directly. For a studio that already has a visual language, that is a meaningful capability.

The honest tradeoff is cost and setup. Scenario plans start around $15/mo and scale significantly for higher-volume or enterprise usage (as of this writing). Training a style model takes time and a sample set, so the payoff is highest if your project already has a defined look and you need volume. If you are still in early exploration or working solo, the setup cost is hard to justify.

Scenario does not offer general-purpose presets, mask editing, or local background removal. Its strength is fidelity to a trained style, not Swiss-army-knife convenience. The deeper comparison is in AgentBrush vs Scenario.

3. MeiGen MCP: open-source multi-model bridge

MeiGen MCP is the most ambitious open-source option. It supports a growing roster of models including GPT Image 2, Nanobanana, and ComfyUI (among others), comes with a 1,400+ curated prompt library, and supports batch generation and parallel orchestration. If you want to route jobs through a local ComfyUI install or experiment with a range of backends, this is probably your starting point.

The gaps are real, though. There are no presets, no mask editor, no background removal, and no reference or identity system. MeiGen is built for people who want direct model access and are comfortable handling the surrounding workflow themselves: prompt engineering, file management, and whatever consistency tricks they need. That is a legitimate use case, but it is a different audience than someone who wants images to "just work" from their agent.

Self-hosting also means you manage updates, credentials, and whatever the underlying model APIs require. That is fine if it is your thing, and genuinely worth noting if it is not.

4. imagegen-mcp (spartanz51): thin but functional

imagegen-mcp is a lightweight open-source server wired to OpenAI's image models (gpt-image-1 and DALL-E variants). It handles text-to-image generation and has partial support for masked edits. It is the simplest way to get OpenAI image generation over MCP if you are using your own API key and do not need anything beyond basic generation.

What it does not have: presets, background removal, reference images, or a mask editing UI. It is a thin wrapper, which is either exactly right or exactly wrong depending on your project. For hackers who want to build something custom on top of OpenAI image generation without any opinions in the way, it is a reasonable base. For anyone who wants a finished tool, it is a starting point that you would need to extend yourself.

5. GongRzhe Flux bridge: Replicate-hosted Flux

This server routes calls to Flux running on Replicate. If you specifically want Flux's aesthetic (which is genuinely different from gpt-image-2, particularly for photorealistic renders with a certain cinematic quality), it is the cleanest way to reach it over MCP.

The scope is limited to generation: batch support and aspect-ratio control are in, but editing, presets, background removal, and reference consistency are not. Replicate's pay-per-use model also means the cost structure is different from a subscription, which can work out cheaper or more expensive depending on volume. Worth knowing about; probably not your primary tool unless Flux is specifically what you need.

6. PixelLab MCP and Retro Diffusion: pixel-art niches

Both are intentionally narrow. PixelLab MCP connects to PixelLab's pixel-art generation service, with tools for characters, animations, tilesets, and isometric tiles. Retro Diffusion is a family of specialized pixel-art models (available through Scenario's platform as of this writing) that outputs transparent-background sprites and animations, which is directly useful for 2D game devs working in pixel-art style.

If pixel art is your whole output type and you want a specialized model trained for it, either of these is worth evaluating. If you need pixel art as one of several styles, AgentBrush's pixel_art preset covers the common cases and keeps everything in one server.

Where the open-source bridges generally fall short

The pattern across MeiGen, imagegen-mcp, and the Flux bridge is the same: they are API wrappers, not finished tools. That is not a criticism of the quality; it is an architectural choice. The author exposes the model endpoint over MCP and leaves the rest to you. That is genuinely useful if "the rest" is something you want to own.

What it means in practice: no presets (every call needs a full prompt), no background removal (you wire in your own tool), no mask editor (you build one or skip editing), no reference consistency system (you solve it yourself with prompt engineering or separate tooling). Each of those is a solvable problem, but the solving is on you.

The open-source bridges are a good fit for builders who want maximum control and are comfortable running their own infrastructure. They are a poor fit if you want a clean "ask your agent for an image" experience without glue work.

How to choose

Start with your constraint:

• Most projects, shipping fast: AgentBrush. Presets, editing, background removal, reference images, and it works with any MCP client out of the box.
• Game studio with a trained art style: Scenario. The trained-model strength is real; just budget for setup and ongoing cost.
• Maximum model flexibility, okay with self-hosting: MeiGen MCP. Multi-model support is the strongest in class among open-source options.
• Direct OpenAI image access, building your own wrapper: imagegen-mcp or the Replicate Flux bridge, depending on which model you want.
• Pixel-art sprites specifically: PixelLab MCP if you want animated characters and tilesets; AgentBrush pixel_art preset if you want pixel art as one style among several.

If you are new to running images inside an agent, the fastest path is generating images in Claude and Cursor, then the AgentBrush install guide once you decide on the tool.

Where this breaks down

MCP server ecosystems move fast. The landscape described here reflects the state of things as of mid-2026. Open-source projects add features, pricing changes, and new servers appear regularly. Confirm capabilities against a project's current README before building a workflow around it.

Open-source does not mean free end-to-end. imagegen-mcp and MeiGen MCP are free to run, but you still pay the underlying model providers (OpenAI, Replicate, etc.) on your own account. For high-volume usage the per-token API costs can exceed a flat subscription.

"MCP-compatible" varies. Not every client implements the full MCP spec equally. AgentBrush is tested against Claude Code, Cursor, Codex CLI, and Gemini CLI. For other clients, verify the integration works before committing to a workflow.

FAQ

What is an image generation MCP server? An MCP server that exposes image-generation capabilities as tools your agent can call. Instead of switching to a web app and pasting the result back, your agent (Claude, Cursor, Codex CLI, etc.) calls the server directly and the image lands in your project. The Model Context Protocol is the standard that makes the tool call possible.

Which MCP server is best for generating images in Claude or Cursor? AgentBrush is the most complete option for general use, with presets, background removal, mask editing, and reference-image consistency. If you specifically need a trained game-art style, Scenario is the better fit. Open-source wrappers like MeiGen MCP and imagegen-mcp are worth considering if you want raw model access and are comfortable with self-hosting.

Do I need my own API key to use these servers? It depends. AgentBrush and Scenario are subscription-based: you pay the platform and they handle the model access. MeiGen MCP and imagegen-mcp require your own OpenAI (or other provider) API key and you pay the model provider directly. The Flux bridge requires a Replicate account.

Can any of these generate a truly transparent PNG? Most image models, including gpt-image-2, do not reliably output a real alpha channel at generation time. AgentBrush handles this in a two-step flow: generate on a clean background, then run local background removal (0 tokens, always returns a real PNG). The open-source wrappers leave background removal as your own problem to solve.

Ready to run your first image from inside your agent? Connect AgentBrush and paste a prompt. It takes about two minutes to set up and your first image lands in your project folder without leaving your editor.