Stability AI MCP Server
Haven't heard about MCP yet? The easiest way to keep up-to-date is to read our weekly newsletter at PulseMCP.
This is an MCP (Model Context Protocol) Server integrating MCP Clients with 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. You will need to sign up for an API Key from stability.ai to get started.
They provide 25 credits for free. Afterward, pay-as-you-go 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, and a teaser here:
Table of Contents
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_DIRECTORYand 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
Generate an image of a catGenerate a photorealistic image of a cat in a cyberpunk city, neon lights reflecting off its fur, 16:9 aspect ratioGenerate 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
Generate an image of a woman with cybernetic wolf ears using the SD3.5 model, with the "neon-punk" style presetGenerate an image of a futuristic city using the SD3.5 Large Turbo model, with aspect ratio 16:9Generate 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
Remove the background from the image I just generatedRemove the background from product-photo.jpg to prepare it for my e-commerce siteRemove the background from group-photo.png so I can composite it with another image
Outpaint (Uncrop)
Extend vacation-photo.jpg 500 pixels to the right to include more of the beachExtend family-portrait.png 300 pixels up to show more of the mountains, and 200 pixels right to include more landscapeExtend artwork.png in all directions to create a wider fantasy forest scene that matches the original environment
Search and Replace
In my last image, replace the red car with a blue carIn portrait.png, replace the plain background with a sunset over mountainsIn landscape.jpg, replace the modern buildings with victorian-era architecture while maintaining the same atmosphere
Upscale
Upscale profile-pic.jpg for better resolutionUpscale product-photo.png to make it print-ready
And then, if the output still isn't good enough, you can upscale it again:
Try again with better quality
Control Sketch
Transform sketch.png into a colored illustration for a children's bookConvert wireframe.jpg into a detailed 3D render for a modern architectural visualization
Control Style
Generate an image in the style of the reference image
Control Structure
Generate an image while maintaining the structure of the reference image
Replace Background and Relight
Replace the background of the image I just generated with a sunset over mountains
Search and Recolor
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 | Required | Default Value | Example |
|---|---|---|---|---|
STABILITY_AI_API_KEY | Your Stability AI API key. Get one at platform.stability.ai | Y | N/A | sk-1234567890 |
IMAGE_STORAGE_DIRECTORY | Directory where generated images will be saved | N | /tmp/tadasant-mcp-server-stability-ai OR C:\\Windows\\Temp\\mcp-server-stability-ai | /Users/admin/Downloads/stability-ai-images (Mac OS/Linux), C:\\Users\\Admin\\Downloads\\stability-ai-images (Windows) |
SAVE_METADATA | Save metadata for successful image generations | N | true | true or false |
SAVE_METADATA_FAILED | Save metadata for failed image generations | N | true | true or false |
GCS_PROJECT_ID | Google Cloud Project ID for storing images | N (Y if using SSE) | N/A | your-project-id |
GCS_CLIENT_EMAIL | Google Cloud Service Account client email for storing images | N (Y if using SSE) | N/A | your-service-account@project.iam.gserviceaccount.com |
GCS_PRIVATE_KEY | Google Cloud Service Account private key for storing images | N (Y if using SSE) | N/A | -----BEGIN PRIVATE KEY-----\nYourKeyHere\n-----END PRIVATE KEY-----\n |
GCS_BUCKET_NAME | Google Cloud Storage bucket name for storing images | N (Y if using SSE) | N/A | your-bucket-name |
Claude Desktop
If you prefer a video tutorial, here's a quick one.
Create a folder directory somewhere on your machine to store generated/modified images. Some options:
/Users/<username>/Downloads/stability-ai-images/Users/<username>/Library/Application Support/Claude/mcp-server-stability-ai/images
And make sure you have an API key from Stability AI.
Then proceed to your preferred method of configuring the server below. If this is your first time using MCP Servers, you'll want to make sure you have the Claude Desktop application and follow the official MCP setup instructions.
Manual Setup
You're going to need Node working on your machine so you can run npx commands in your terminal. If you don't have Node, you can install it from nodejs.org.
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Modify your claude_desktop_config.json file to add the following:
{
"mcpServers": {
"stability-ai": {
"command": "npx",
"args": [
"-y",
"mcp-server-stability-ai"
],
"env": {
"STABILITY_AI_API_KEY": "sk-1234567890"
}
}
}
}
Restart Claude Desktop and you should be ready to go:
Installing via Smithery
To install for Claude Desktop automatically via Smithery:
npx @smithery/cli install mcp-server-stability-ai --client claude
SSE Mode
This server has the option to run in SSE mode by starting it with the following command:
npx mcp-server-stability-ai -y --sse
This mode is useful if you intend to deploy this server for third party usage over HTTP.
You will need to set the GCS_PROJECT_ID, GCS_CLIENT_EMAIL, GCS_BUCKET_NAME, and GCS_PRIVATE_KEY environment variables, because the server will store image files in Google Cloud Storage instead of its local filesystem.
Note that the scheme for multitenancy is very naive and insecure: it uses the requestor's IP address to segment the GCS prefixes used to the store the images, and makes all images publicly accessible in order to communicate them back to the MCP client. So in theory, if someone knows your IP address and then name(s) of files you generated, they could access your images by guessing the URL.
Roadmap
Recently completed:
- ✅ Added support for Stable Diffusion 3.5 models
- ✅ Added support for Stable Image Ultra
- ✅ Added metadata logging for image generation requests
These are coming soon; but PR's are welcome! See CONTRIBUTING.md.
inpaint- A more precise version ofsearch-and-xfunctionalities, requires mananging a mask to define to area to replace.- Base image manipulation (crop, rotate, resize, etc.): likely as its own MCP server
- Ability to inpaint one image into another image. Doesn't seem possible with Stability API; will probably want another MCP server hitting a different API to accomplish this.
- MCP client custom-made for image manipulation
Contributing
External contributions are welcome! See CONTRIBUTING.md for more details.
Also please feel free to raise issues or feature requests; love seeing how people are using this and how it could be made better.