ImageSorcery MCP: local image editing and recognition

Many agent image tools send pictures to a cloud model to “look” at them, which feels wrong for screenshots, IDs or internal documents. ImageSorcery MCP stays local: OpenCV for cropping, resizing, drawing and blurring, Ultralytics models for object detection, EasyOCR for text, all running on your machine. The repo had about 330 stars as of 2026-10-06.

What it does

• Basic edits: crop, resize, rotate, change_color (grayscale or sepia, say) and overlay for logos and watermarks.
• Annotation and redaction: draw_rectangles, draw_arrows and draw_texts mark up images, blur hides a region, and fill paints an area or makes it transparent.
• Recognition: detect finds objects with pretrained models and can return segmentation masks, find locates objects from a text description (“the dog in this photo”), and ocr extracts text.
• Chained tasks: the model strings tools together, for example find the cat, then crop so it’s centered; the built-in remove-background prompt walks through cutting out a subject.

Who it’s for

• People writing docs and tutorials who want arrows, boxes and redactions added to a batch of screenshots.
• Anyone sorting photos who wants them filed by content, such as “photos with pets”.
• People handling scans and forms who want text or form fields extracted before tidying them up.

Setup

You need Python 3.10+ and pipx. Run the post-install step once after installing; it downloads the detection models:

pipx install imagesorcery-mcp
imagesorcery-mcp --post-install

Then add "imagesorcery-mcp": {"command": "imagesorcery-mcp", "timeout": 100} to your client config.

Our take

Most image servers call a generation model. ImageSorcery handles the everyday jobs: edit, annotate, recognize, offline. That’s why it’s here. Some caveats. File paths are unrestricted by default, so the agent can read and write images anywhere on your machine; set IMAGESORCERY_AVAILABLE_PATHS to fence it in. The first install downloads sizable models, and slim Linux or Docker images may lack OpenCV’s system libraries. Anonymous telemetry is off by default, but the README’s one-shot install prompt for Cline asks the agent to turn it on after asking you, so watch for that. The last commit was in May 2026, so updates aren’t frequent. MIT-licensed.