ImageSorcery MCP
Local computer-vision image recognition and editing tools for AI assistants — crop, detect, OCR, remove background, all on-device.
Add to your client
Copy the config for your MCP client and paste it into its config file.
pipx install imagesorcery-mcpPaste into ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"imagesorcery-mcp": {
"command": "imagesorcery-mcp",
"args": []
}
}
}Step-by-step guides: Add to Claude Desktop · Add to Cursor · Add to Windsurf
Before you start
- Python 3.10 or higher
- pipx (recommended) for installation and virtual-environment management
- System libraries required by OpenCV: ffmpeg, libsm6, libxext6, libgl1-mesa-glx (typically bundled with OpenCV; may be missing in containers like Docker)
- An MCP client (Claude.app, Cline, Cursor, or another)
- Run 'imagesorcery-mcp --post-install' after install to download default YOLO models and the CLIP package
About ImageSorcery MCP
ImageSorcery MCP exposes a suite of OpenCV- and ML-powered image tools over the Model Context Protocol so AI assistants (Claude, Cursor, Cline, and other MCP clients) can recognize and edit images entirely on the local machine. It bundles drawing primitives, geometric transforms, color/format changes, object detection via Ultralytics YOLO models, text-prompted object finding via CLIP, OCR via EasyOCR, and background-removal workflows. A post-install step downloads default models and the CLIP package. The server runs over stdio by default and can also serve Streamable HTTP and SSE endpoints for web deployments.
Tools & capabilities (17)
blurBlurs specified rectangular or polygonal areas of an image using OpenCV; can invert the provided areas (e.g. to blur the background).
change_colorChanges the color palette of an image (e.g. convert to sepia).
configView and update ImageSorcery MCP configuration settings (e.g. default detection confidence, telemetry).
cropCrops an image using OpenCV's NumPy slicing approach between two coordinates.
detectDetects objects in an image using Ultralytics models; can return segmentation masks (as PNG files) or polygons.
draw_arrowsDraws arrows on an image using OpenCV.
draw_circlesDraws circles on an image using OpenCV.
draw_linesDraws lines on an image using OpenCV.
draw_rectanglesDraws rectangles (outlined or filled) on an image using OpenCV.
draw_textsDraws text at specified positions on an image using OpenCV.
fillFills rectangular, polygonal, or mask-based areas with a color and opacity, or makes them transparent; can invert areas (e.g. to remove background).
findFinds objects in an image based on a text description; can return segmentation masks (as PNG files) or polygons.
get_metainfoGets metadata information about an image file.
ocrPerforms Optical Character Recognition (OCR) on an image using EasyOCR.
overlayOverlays one image on top of another, handling transparency (e.g. placing a logo).
resizeResizes an image to target dimensions using OpenCV.
rotateRotates an image by a given angle using imutils.rotate_bound.
When to use it
- Detect and crop objects (e.g. find a cat and center it) automatically
- Batch-organize photos by content, e.g. copy images containing pets to a separate folder
- Extract text from documents and screenshots with local OCR
- Remove or blur backgrounds while keeping a detected subject
- Add watermarks, logos, text, and shapes to images
- Enumerate and describe UI/form fields in screenshots using a detection model
Security notes
Runs fully locally — images and data stay on your machine. By default no file paths are restricted; set IMAGESORCERY_AVAILABLE_PATHS to one or more allowed directories to limit tools to specific folders (path-list separator ':' on Linux/macOS, ';' on Windows; commas also accepted). Optional anonymous telemetry is disabled by default and must be explicitly opted into; it never collects image data, file paths, IPs, or other PII.
ImageSorcery MCP FAQ
Are my images sent to any external server?
No. ImageSorcery is designed to run entirely locally; your images and data stay on your machine. Optional anonymous telemetry is disabled by default and never includes image data or file paths.
What does the post-install step do?
Running 'imagesorcery-mcp --post-install' creates a config.toml, creates a models directory with model_descriptions.json, downloads default YOLO models for the detect tool, attempts to install the CLIP Python package from Ultralytics' GitHub (needed for the find tool's text prompts), and downloads the CLIP model file.
Can I restrict which directories the tools can access?
Yes. Set the IMAGESORCERY_AVAILABLE_PATHS environment variable to one or more allowed directories. When set, every tool argument named 'path' or ending with '_path' must resolve inside an allowed directory.
Does it support HTTP transports?
Yes. Besides the default stdio mode, you can run it with --transport=streamable-http or --transport=sse, with configurable --host, --port, and --path, then point your MCP client at the HTTP endpoint (default http://127.0.0.1:8000/mcp).
Alternatives to ImageSorcery MCP
Compare all alternatives →
Control Blender from Claude and other LLMs for prompt-assisted 3D modeling, scene creation, and asset generation.
Popular community server that feeds Figma layout data to coding agents via a Figma API token.
Generate beautiful, modern UI components from natural-language descriptions inside your IDE.
Compare ImageSorcery MCP with: