Haven't heard about MCP yet? The easiest way to keep up-to-date is to read our [weekly newsletter at PulseMCP](https://www.pulsemcp.com/). --- This is an MCP ([Model Context Protocol](https://modelcontextprotocol.io/)) Server integrating MCP Clients with [Stability AI](https://stability.ai/)'s latest & greatest Stable Diffusion image manipulation functionalities: generate, edit, upscale, and more. Stability AI is a leading AI model provider and this server connects directly to their [hosted REST API](https://platform.stability.ai/docs/api-reference). You will need to sign up for an [API Key from stability.ai](https://platform.stability.ai/account/keys) to get started. They provide 25 credits for free. Afterward, [pay-as-you-go pricing](https://platform.stability.ai/pricing) is very reasonable: $0.01/credit, where 3 credits gets you an image generation on their Core model. So 100 high quality images = just $3. This project is NOT officially affiliated with Stability AI. [Demo video](https://youtu.be/7ceSgVC4ZLs), and a teaser here: 
# Table of Contents - [Highlights](#highlights) - [Capabilities](#capabilities) - [Usage Tips](#usage-tips) - [Examples](#examples) - [Setup](#setup) - [Cheatsheet](#cheatsheet) - [Claude Desktop](#claude-desktop) - [Manual Setup](#manual-setup) # Highlights **No learning curve**: This server is designed to use sensible defaults and provide simple, smooth UX for the most common actions related to generating and manipulating image files. You don't need to be technical or understand anything about image models to use it effectively. **Chain manipulations**: You can generate an image, then replace an item within it, then remove the background... all in a single Claude conversation. **Minimal configuration**: All you need to get started is a Stability AI API key. Set environment variables for that and a local directory path to store output images, and you're ready to go. **Leverage the best in class image models**: Stability AI is the leading provider of image models exposed via API. Using this server integrates them into Claude or another MCP client - head and shoulders above an experience like using DALL-E models in ChatGPT. # Capabilities This server is built and tested on macOS with Claude Desktop. It should work with other MCP clients as well. | Tool Name | Description | Estimated Stability API Cost | | -------------------------------- | -------------------------------------------------------------------------------------------------- | ---------------------------- | | `generate-image` | Generate a high quality image of anything based on a provided prompt & other optional parameters. | $0.03 | | `generate-image-sd35` | Generate an image using Stable Diffusion 3.5 models with advanced configuration options. | $0.04-$0.07 | | `remove-background` | Remove the background from an image. | $0.02 | | `outpaint` | Extend an image in any direction while maintaining visual consistency. | $0.04 | | `search-and-replace` | Replace objects or elements in an image by describing what to replace and what to replace it with. | $0.04 | | `upscale-fast` | Enhance image resolution by 4x. | $0.01 | | `upscale-creative` | Enhance image resolution up to 4K. | $0.25 | | `control-sketch` | Translate hand-drawn sketch to production-grade image. | $0.03 | | `control-style` | Generate an image in the style of a reference image. | $0.04 | | `control-structure` | Generate an image while maintaining the structure of a reference image. | $0.03 | | `replace-background-and-relight` | Replace the background of an image and relight it. | $0.08 | | `search-and-recolor` | Search for and recolor objects in an image. | $0.05 | # Usage Tips - All processed images are automatically saved to `IMAGE_STORAGE_DIRECTORY`, opened for preview, and made available as resources - Do _not_ try to copy/paste or upload image files to Claude. Claude does not store images anywhere, so we cannot work with those with the MCP server. They have to be "uploaded" (saved to) the `IMAGE_STORAGE_DIRECTORY` and then they will show up as resources available in the chat. - You can use Prompts that come preloaded instead of writing your own verbiage: # Examples ## Generate an image 1. `Generate an image of a cat` 2. `Generate a photorealistic image of a cat in a cyberpunk city, neon lights reflecting off its fur, 16:9 aspect ratio` 3. `Generate a detailed digital art piece of a cat wearing a space suit floating through a colorful nebula, style preset: digital-art, aspect ratio: 21:9` ## Generate an image with SD3.5 1. `Generate an image of a woman with cybernetic wolf ears using the SD3.5 model, with the "neon-punk" style preset` 2. `Generate an image of a futuristic city using the SD3.5 Large Turbo model, with aspect ratio 16:9` 3. `Generate an image of an astronaut on mars using the SD3.5 Large model, with cfg scale 7.5, "analog-film" style preset, and seed 42` ## Remove background 1. `Remove the background from the image I just generated` 2. `Remove the background from product-photo.jpg to prepare it for my e-commerce site` 3. `Remove the background from group-photo.png so I can composite it with another image` ## Outpaint (Uncrop) 1. `Extend vacation-photo.jpg 500 pixels to the right to include more of the beach` 2. `Extend family-portrait.png 300 pixels up to show more of the mountains, and 200 pixels right to include more landscape` 3. `Extend artwork.png in all directions to create a wider fantasy forest scene that matches the original environment` ## Search and Replace 1. `In my last image, replace the red car with a blue car` 2. `In portrait.png, replace the plain background with a sunset over mountains` 3. `In landscape.jpg, replace the modern buildings with victorian-era architecture while maintaining the same atmosphere` ## Upscale 1. `Upscale profile-pic.jpg for better resolution` 2. `Upscale product-photo.png to make it print-ready` And then, if the output still isn't good enough, you can upscale it again: 1. `Try again with better quality` ## Control Sketch 1. `Transform sketch.png into a colored illustration for a children's book` 2. `Convert wireframe.jpg into a detailed 3D render for a modern architectural visualization` ## Control Style 1. `Generate an image in the style of the reference image` ## Control Structure 1. `Generate an image while maintaining the structure of the reference image` ## Replace Background and Relight 1. `Replace the background of the image I just generated with a sunset over mountains` ## Search and Recolor 1. `In my last image, make the red car be blue instead` # Setup ## Metadata Logging The server can save metadata from image generation requests to help with tracking and troubleshooting. | Environment Variable | Description | Required | Default Value | | ------------------------ | --------------------------------------------------- | -------- | ------------- | | `SAVE_METADATA` | Save metadata for successful image generations | N | `true` | | `SAVE_METADATA_FAILED` | Save metadata for failed image generations | N | `false` | When enabled, a `.txt` file with the same name as the generated image will be created in the same directory. This file contains: - Timestamp of the request - All request parameters (prompt, model, style preset, etc.) - Response information (success status, generation time) This file will also be created for failed requests if `SAVE_METADATA_FAILED` is enabled. ## Cheatsheet | Environment Variable | Description ---