Short answer: The best image generation MCP server depends on where you run Claude and who pays for the model. For a developer on Claude Code or Claude Desktop with an API key, nanobanana-mcp-server (Gemini, including Nano Banana 2) and openai-gpt-image-mcp (gpt-image-1) are the most used open-source wrappers. For Claude on the web, you need a remote server: OpenArt and Higgsfield connect you to their model catalogs, and Avocado AI (our product) connects a brand workspace with images, video and publishing. This roundup compares nine options with facts taken from each project's own page.
Updated November 2026. Avocado AI publishes this blog and is listed as option 9. Every competitor fact below comes from the linked README or product page; where a page does not state a price, we say so.
How to read this list
An MCP server for image generation does one thing: it exposes a generate tool (and often an edit tool) that Claude, Cursor, Codex or any MCP client can call. The differences that matter in practice:
- Local or remote. Local servers run on your machine and only work in Claude Desktop and Claude Code. Remote servers run on a URL and also work in claude.ai, the mobile apps and Cowork.
- Who pays for the model. Open-source wrappers use your own API key; platform servers use their credits.
- Where the image goes. A folder on disk, a chat message, or a library your team can open later.
- Video. Most image servers stop at images.
Comparison table
| Server | Hosting | Auth | Models | Free or paid | Video |
|---|---|---|---|---|---|
| nanobanana-mcp-server | Local (uvx or pip) | GEMINI_API_KEY or Vertex ADC | Nano Banana 2 (default), Gemini 3 Pro Image, Gemini 2.5 Flash Image | Free code, you pay Google per image | No |
| openai-gpt-image-mcp | Local (Node) | OPENAI_API_KEY or Azure | gpt-image-1 | Free code, you pay OpenAI per image | No |
| img-gen-mcp | Local (npx) or opt-in HTTP | Provider key in your env | About 22 OpenRouter models plus OpenAI, Gemini and local backends | Free code, you pay the provider | Not mentioned |
| universal-image-generator-mcp | Local (uvx) | GEMINI_API_KEY, ZHIPU_API_KEY or DASHSCOPE_API_KEY | Google Imagen and Gemini, CogView-4, WanX-2.1 | Free code, you pay the provider | No |
| imagine-mcp | Local (uvx or Docker), optional HTTP | GEMINI, OPENAI or XAI key | Gemini image and Veo 3.1, gpt-image, Grok Imagine and Grok Imagine Video | Free code, archived September 2026 | Yes |
| drzippie/image-generator | Local (Go) | WaveSpeed API key in config.yaml | WaveSpeed z-image turbo and base | Free code, you pay WaveSpeed | No |
| OpenArt MCP | Remote (Streamable HTTP) | OAuth with an OpenArt account | 8 image and 8 video models listed | OpenArt credits, same as its web app | Yes |
| Higgsfield MCP | Remote (custom connector) | Sign in with a Higgsfield account | "15+" models named on the page | Free renders to start, then credits; no prices on the page | Page covers images; Kling O1 is listed among models |
| Avocado AI MCP | Remote (Streamable HTTP) | OAuth in Claude, API key elsewhere | GPT Image 2.5, Nano Banana 2, Seedream 5.0 Pro, Flux 2 and more; Seedance 2.5, Kling 3.0, Veo 3.1 for video | Paid plans from 19.99 EUR a month, 100 credits; no free plan | Yes |
1. nanobanana-mcp-server (Gemini, local)
Verdict: the best open-source pick if you already have a Gemini API key.
zhongweili/nanobanana-mcp-server describes itself as an "AI image generation MCP server powered by Google Gemini, with smart model selection and 4K output." It ships three Gemini image models with automatic selection: Nano Banana 2 (gemini-3.1-flash-image-preview) as the default with up to 4K output, Gemini 3 Pro Image for demanding compositions, and Gemini 2.5 Flash Image as a legacy option up to 1024px. Tools named on the page are generate_image, edit_image and upload_file, plus prompt templates and file management through the Gemini Files API.
It runs in stdio mode; install with uvx nanobanana-mcp-server@latest or pip install nanobanana-mcp-server. Authentication is a GEMINI_API_KEY or Vertex AI through Google Cloud Application Default Credentials. MIT license, 404 stars and 114 forks at the time of writing.
Good for: developers who want Nano Banana 2 inside Claude Code or Cursor without a platform account. Not for: Claude on the web (local only), teams that need shared assets, or anyone who wants video.
2. openai-gpt-image-mcp (gpt-image-1, local)
Verdict: the straightforward OpenAI wrapper, with editing and Azure support.
SureScaleAI/openai-gpt-image-mcp is "a Model Context Protocol (MCP) tool server for OpenAI's GPT-4o/gpt-image-1 image generation and editing APIs." Two tools: create-image (size, quality, background, up to 10 images per call) and edit-image (prompt plus optional mask, file path or base64, for inpainting, outpainting and compositing). Images can be saved to disk or returned as base64; responses over 1 MB switch to file output automatically.
Setup is clone, yarn install, yarn build, then point your client at node dist/index.js with OPENAI_API_KEY in env. The README notes that the OpenAI account must be a verified organization and that image API access can take 15 to 20 minutes to activate after verification. Azure OpenAI is supported through AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT and OPENAI_API_VERSION. MIT license, 119 stars.
Good for: teams standardised on OpenAI or Azure. Not for: anyone who wants a model choice, or who runs Claude in the browser.
3. img-gen-mcp (multi-provider, local-first)
Verdict: the most feature-rich open-source option, at the cost of complexity.
jgkme/img-gen-mcp calls itself "a local-first MCP server for image generation, editing, async jobs, batch runs, background removal, optimization, final asset delivery, and optional ACP-style automation." Providers are OpenRouter (default), Kilo Gateway, OpenAI and Gemini, plus local backends such as OpenAI-compatible endpoints, MLX, ComfyUI, Draw Things and llama.cpp servers. The OpenRouter list includes about 22 models, among them microsoft/mai-image-2.5, openai/gpt-image-1, google/gemini-2.5-flash-image, recraft/recraft-v4.1-vector and black-forest-labs/flux.2-pro.
It runs on Node: npx -y img-gen-mcp, a local checkout, or an opt-in Streamable HTTP mode with npx -y img-gen-mcp http. Background removal runs locally (rmbg, imgly, withoutBG) and web optimisation uses sharp. MIT license, 3 stars, 75 commits, version 0.11.0 on the page.
Good for: developers who want batch runs and background removal from one server. Not for: non-technical users; the README is long and the key setup is up to you.
4. universal-image-generator-mcp (Google, Zhipu, Alibaba; local)
Verdict: a small, clean multi-provider server with a Chinese-model angle.
ECNU3D/universal-image-generator-mcp is a fork and rewrite of an earlier Gemini image server. It supports Google (Imagen and Gemini), ZhipuAI (CogView-4) and Alibaba Bailian (WanX-2.1). All three generate from text; image transformation works with Google (Gemini models only) and Bailian, not Zhipu. Output is saved to a configurable local directory.
It runs via uvx with the package name universal-image-generator-mcp, configured with IMAGE_PROVIDER, GOOGLE_MODEL (gemini or imagen), OUTPUT_IMAGE_PATH and the matching key: GEMINI_API_KEY, ZHIPU_API_KEY or DASHSCOPE_API_KEY. MIT license, 2 stars.
Good for: anyone who needs CogView or WanX next to Gemini. Not for: editing-heavy work or video.
5. imagine-mcp (Gemini, OpenAI, Grok; archived)
Verdict: the only open-source option here with video, but it is no longer maintained.
n24q02m/imagine-mcp covers image and video understanding and generation across Gemini (Imagen and Veo 3.1), OpenAI (gpt-image) and Grok (Grok Imagine and Grok Imagine Video), with a poor or rich tier per provider. Tools: understand, generate, config and help. It runs with uvx imagine-mcp or Docker, stdio by default or an HTTP daemon with imagine-mcp --http. Keys: GEMINI_API_KEY, OPENAI_API_KEY, XAI_API_KEY (at least one). Apache-2.0, 4 stars.
The repository was archived on September 13, 2026, with a notice that it "is no longer maintained" and a recommendation to use each provider's native API instead. We list it because it still appears in search results and the install still works, but do not build on it.
6. drzippie/image-generator (WaveSpeed, local)
Verdict: a minimal Go CLI with an MCP mode, for one provider.
drzippie/image-generator is a CLI tool and MCP server for the WaveSpeed.ai API. Default model is wavespeed-ai/z-image/turbo (faster and cheaper per the README); z-image/base is the higher-quality option. Width and height default to 1024x1024, with PNG, JPEG and WebP output. The MCP server is built from cmd/mcp-server and registered in Claude Desktop's config file. Requires Go 1.21 or higher and a WaveSpeed API key in config.yaml. MIT license, 0 stars.
Good for: a WaveSpeed user who wants a tiny binary. Not for: anyone else.
7. OpenArt MCP (remote, OAuth)
Verdict: the broadest hosted catalog for Claude web users who want models, not a workspace.
OpenArt's MCP page gives the endpoint https://mcp.openart.ai/mcp over Streamable HTTP with OAuth sign-in and no API key. Capabilities listed: image and video generation from text, image-to-video, external image references, a Smart Shot multi-shot video planner, and saving to your OpenArt library. The page lists eight image models (Nano Banana 2.1, Nano Banana Pro, GPT Image 2.5, GPT Image 2, Seedream 5 Pro, Wan 2.7 Image, Ideogram 4.5, Grok Imagine Image 2.0) and eight video models (Seedance 2.5, Seedance 2.0, Wan 3.0, Kling 3 Omni, Veo 3.1, MiniMax H3, Gemini Omni 1.1 Flash, PixVerse V6). Character is listed as roadmap.
Setup in Claude is Customize, Connectors, Add custom connector, paste the URL, Connect, sign in. MCP generations use the same credits as openart.ai, priced by model and resolution; the page does not state whether a free plan includes MCP access. OpenArt says Claude Code and Codex users are better served by its CLI.
Good for: trying many models from one connector. Not for: brand consistency across a campaign; there is no product or brand memory described on the page.
8. Higgsfield MCP (remote, account sign-in)
Verdict: a strong creative catalog with character consistency, light on published detail.
Higgsfield's Claude AI image generator page describes adding Higgsfield as a custom connector in Claude (Settings, Connectors, add a custom connector named Higgsfield with the MCP URL shown on the page, then Connect and sign in). It names "15+" models including Nano Banana Pro, Higgsfield Soul, Seedream, FLUX, GPT Image, Kling O1 and Wan, and says Claude can pick a model automatically or run several in parallel. Soul ID trains a character from a few photos for reuse. Output up to 4K in 1:1, 9:16, 16:9, 4:5 and 3:4; paid plans are watermark-free.
On price, the page says Higgsfield "offers free renders to start" and that paid usage runs on credits priced by model and resolution, with no specific prices given. For Claude Code the page points to a CLI instead.
Good for: stylised campaign imagery and consistent characters. Not for: anyone who needs published pricing before connecting, or video and audio in the same tool.
9. Avocado AI MCP (remote, workspace)
Verdict: our pick for marketing teams; a remote server where every output lands in a brand workspace, with video, audio and publishing in the same tool.
Disclosure: this is our product, so weigh this section accordingly. The Avocado MCP server lives at https://mcp.avocadoai.co/api/mcp over Streamable HTTP. Claude signs in with OAuth; Claude Code, Codex, Cursor, Windsurf, Hermes and self-built agents use an Avocado API key as a Bearer token. It exposes 60+ tools: generate_image, edit_image, generate_video, generate_music, generate_sfx and generate_speech; brand tools (extract_brand, apply_brand_layer, extract_product, remember); Flow tools for repeatable campaigns; publish and schedule tools with the Connectors add-on; and account tools including a free models_list that needs no key.
Images default to GPT Image 2.5, with Nano Banana 2, Seedream 5.0 Pro, Soul 2, Recraft v4.1, Krea v2, Luma Uni-1 Max, Ideogram v3 and Flux 2 by name. Video: Seedance 2.5 (up to 30 seconds), Seedance 2.0, Kling 3.0 (up to 4K), Veo 3.1, MiniMax H3 and others. Every result is saved to your Avocado workspace and gallery, Claude can read your Brand DNA (products, references, tone and rules) before it prompts, and the agent sees the credit cost before it generates. Full commercial rights and no watermark on every plan.
Pricing, in euros: Intro 19.99 a month for 100 credits, Starter 39 for 300, Growth 99 for 800, Pro 249 for 2,000, with 20% off yearly. There is no free plan and no free trial.
Good for: ecommerce and DTC brands and agencies that need product images, video, UGC and voice to match across a campaign, and that want to post from the same chat. Not for: developers who want a local, offline server, anyone who only needs a few free renders, or users who want to bring their own API keys. Setup details and prompts are on the MCP page; a vendor-neutral walkthrough is in how to add an MCP server to Claude.
Which image generation MCP server should you choose?
| If you are... | Pick | Why |
|---|---|---|
| A developer on Claude Code with a Gemini key | nanobanana-mcp-server | Nano Banana 2 by default, 4K, editing, active repo |
| A developer standardised on OpenAI or Azure | openai-gpt-image-mcp | Simple, editing with masks, Azure variables |
| A tinkerer who wants batch runs and background removal | img-gen-mcp | 22+ OpenRouter models, local backends |
| A Claude web user who wants to try many models | OpenArt MCP | 16 models listed, OAuth, no key |
| A creator chasing a stylised look with a consistent character | Higgsfield MCP | Soul ID, 15+ models, free renders to start |
| A brand or agency shipping campaigns | Avocado AI MCP | Workspace, Brand DNA, video and audio, publishing, costs shown up front |
How we compiled this list
We took the servers that appear in Google's top results for "image generation MCP server" and "best mcp servers for image generation" in November 2026, opened each GitHub README or product page, and recorded only what the page states: models, auth, transport, license, stars and pricing. We did not benchmark output quality, because every open-source server here calls the same hosted models you can test directly, and quality is a property of the model, not the wrapper. Star counts move; treat them as a rough signal of maintenance, not of quality.
FAQ
What is an image generation MCP server?
A Model Context Protocol server that exposes image tools (generate, edit, sometimes upscale or video) to an AI client such as Claude, Cursor or Codex. The client shows the tools to the model, the model calls one when a request needs it, and the server runs the image model and returns the result.
Do I need an API key to use an image MCP server with Claude?
For open-source local servers, yes: a Gemini, OpenAI, xAI, Zhipu or other provider key, billed by that provider. For remote platform servers (OpenArt, Higgsfield, Avocado), you sign in with the platform account in Claude; Avocado uses an API key only in non-Claude clients.
Can I use these in Claude on the web?
Only the remote ones: OpenArt, Higgsfield and Avocado. Local servers (everything with uvx, npx, node or Go in its install) work in Claude Desktop and Claude Code, not in claude.ai, because remote connectors are brokered through Anthropic's cloud and cannot reach your machine.
Which MCP server supports video as well as images?
Among the nine: imagine-mcp (archived), OpenArt, Higgsfield (Kling O1 is listed, the page is image-focused) and Avocado (Seedance 2.5, Kling 3.0, Veo 3.1, MiniMax H3).
Is there an official Anthropic image generation MCP server?
Not that we found. Anthropic's help center says Claude does not generate photos or illustrations natively, and its published skills cover documents, not images. Image generation comes from third-party servers like the ones above.
Which is the cheapest?
For a developer, an open-source wrapper plus your own key, because the code is free and you pay the provider's per-image rate. For a team, compare the credit price per image on each platform against the time saved by having assets in one library; the platform pages above state credits but mostly not per-image prices, and Avocado's plans start at 19.99 EUR for 100 credits.
Next step
Developer with a key: install nanobanana-mcp-server or openai-gpt-image-mcp tonight. Marketer in Claude web: add one remote connector and start a new chat. If the job is on-brand product images and video for a store, connect Avocado and brief Claude the way you would brief a designer; the AI image generator shows the same models in the web app.