ImageSorcery MCP

Detects objects in images using YOLO (You Only Look Once) models running locally via the FastMCP server, returning structured bounding box coordinates, class labels, and confidence scores without sending image data to external APIs. The system manages model lifecycle through a post-installation script that automatically downloads YOLO weights and caches them in the models/ directory, enabling offline operation after initial setup.

Performs zero-shot image classification and semantic search using CLIP (Contrastive Language-Image Pre-training) models that encode both images and text into a shared embedding space, enabling AI assistants to classify images against arbitrary text labels without retraining. The system uses cosine similarity between image and text embeddings to rank matches, with model weights automatically downloaded via download_clip.py during setup.

Composites multiple images together using alpha blending and layer operations through OpenCV's addWeighted and bitwise operations, enabling AI assistants to combine images, apply watermarks, or create composite visualizations. The capability supports configurable opacity, blending modes, and positioning of overlay images.

Draws text, rectangles, circles, lines, and arrows on images using OpenCV's drawing functions (putText, rectangle, circle, line, arrowedLine), enabling AI assistants to annotate detection results, create visualizations, or mark regions of interest. The capability supports configurable colors, line widths, and font properties for flexible annotation styling.

Exposes image processing operations as MCP tools with standardized schema-based parameter validation, enabling AI clients (Claude, Cursor, Cline) to discover, invoke, and chain image processing operations through the Model Control Protocol. The FastMCP framework handles tool registration, parameter marshaling, and error handling through a middleware stack that validates inputs against JSON schemas.

Automatically downloads, caches, and manages computer vision model weights (YOLO, CLIP, EasyOCR) through post-installation scripts (post_install.py, download_models.py, download_clip.py) that provision models into a models/ directory, enabling zero-configuration operation after setup. The system tracks model metadata and provides resource listings through the models://list resource.

Defines multi-step image processing workflows (e.g., remove-background) as MCP prompts that orchestrate multiple tools in sequence, enabling AI assistants to execute complex operations through natural language instructions that are expanded into tool invocation chains. The system uses prompt templates to guide AI reasoning and tool selection.

Provides a configuration system (config.py) that manages runtime parameters for image processing operations, model selection, and server behavior through environment variables and configuration files. The system exposes a config tool through MCP that allows AI assistants to query and modify settings at runtime without restarting the server.

Extracts text from images using EasyOCR, a multi-language optical character recognition library that runs locally within the MCP server, returning recognized text with bounding boxes and confidence scores. The system supports 80+ languages and handles rotated/skewed text through preprocessing, with model weights cached after initial download.

Extracts comprehensive metadata from images including dimensions, color space, EXIF data, file size, and format information through OpenCV and PIL/Pillow libraries integrated into the MCP server. This capability provides structured analysis of image properties without modification, enabling AI assistants to understand image characteristics before applying transformations.

Crops images to specified rectangular regions using pixel-level coordinate inputs (x, y, width, height) through OpenCV's array slicing, enabling AI assistants to extract specific areas of interest identified by detection or analysis tools. The capability preserves image quality and supports both absolute coordinates and relative positioning.

Applies Gaussian blur filters to images using OpenCV's GaussianBlur function with configurable kernel size and sigma parameters, enabling AI assistants to reduce noise, create visual effects, or prepare images for downstream analysis. The implementation supports both standard Gaussian blur and bilateral filtering for edge-preserving smoothing.

Fills connected regions of similar color with a specified color using OpenCV's floodFill function, enabling AI assistants to replace backgrounds, paint regions, or modify specific color areas identified by analysis. The capability supports configurable color tolerance thresholds to control fill boundaries.

Resizes images to specified dimensions using OpenCV's resize function with support for multiple interpolation methods (bilinear, bicubic, Lanczos), enabling AI assistants to scale images for different use cases while controlling quality vs performance tradeoffs. The capability supports both absolute dimensions and aspect-ratio-preserving scaling.

Rotates images by specified angles and applies perspective transformations using OpenCV's warpAffine and warpPerspective functions, enabling AI assistants to correct image orientation, straighten skewed documents, or apply geometric transformations. The capability handles rotation around custom pivot points and supports configurable background fill for rotated areas.

Modifies image colors by adjusting hue, saturation, and brightness in HSV color space using OpenCV's cvtColor and in-place array operations, enabling AI assistants to perform color grading, desaturation, or color-based filtering. The capability converts between RGB and HSV, applies adjustments, and converts back while preserving image structure.