ImageSorcery MCP: Local Image Processing Tools for AI Assistants

What ImageSorcery MCP Provides and Who It Targets

ImageSorcery MCP is designed for developers and AI-assisted workflows where image manipulation tasks arise during work sessions. Instead of writing Python scripts directly with OpenCV or PIL, the user describes the image task in natural language to their AI assistant, and the assistant calls the appropriate tool from the server.

The README gives concrete examples of how this works in practice. A prompt like "copy photos with pets from folder photos to folder pets" triggers the object detection tool to identify pets and the file system tools to move matching images. A more specific prompt like "Find a cat at the photo.jpg and crop the image in a half in height and width to make the cat be centered" chains the detect and crop tools.

The target users are developers working with AI coding assistants who need occasional image operations as part of a broader workflow. It is not a general-purpose desktop image editor. The interface is always the AI assistant's chat or command window, not a graphical image editing UI.

All processing runs locally. The README emphasizes this: every tool operates on local files using local models, without sending images to any server. This makes it suitable for working with proprietary or sensitive images that should not leave the machine.

The 17 Tools and How They Are Invoked

The README documents 17 tools accessible through the MCP server. The tool set covers the most common programmatic image operations:

basic geometry: crop, resize, rotate; drawing: draw_arrows, draw_circles, draw_lines, draw_rectangles, draw_texts; compositing: overlay, fill, blur; color: change_color; utility: get_metainfo; computer vision: detect and find (both using Ultralytics models with segmentation mask or polygon output); text extraction: ocr (using EasyOCR).

The detect tool identifies objects using models from Ultralytics and can return segmentation masks as PNG files or polygon coordinates. The find tool does text-prompted object search: "Find all dogs in my image 'photo.jpg' with a confidence threshold of 0.4".

The ocr tool performs Optical Character Recognition using EasyOCR. The README example specifies language as a parameter: "Extract text from my image 'document.jpg' using OCR with English language".

There is also a config tool for viewing and updating ImageSorcery MCP configuration settings, and a models://list resource that shows which models are available in the models directory.

One bundled prompt, remove-background, guides the AI through a background removal workflow using the object detection and masking tools. The README notes that detailed instructions for each tool, resource, and prompt are in the respective README.md files under /src/imagesorcery_mcp/tools/, /src/imagesorcery_mcp/resources/, and /src/imagesorcery_mcp/prompts/.

Installing ImageSorcery MCP

The recommended installation method uses pipx, which installs the package and manages its virtual environment automatically:

pipx install imagesorcery-mcp

After installation, a post-install step configures the MCP server in the appropriate client configuration file:

imagesorcery-mcp --post-install

For Cline specifically, the README includes an LLM-INSTALL.md that can be pasted directly into the chat. The instructions tell the AI to load MCP documentation, run the two commands above, and modify the MCP configuration file.

The system requirements include Python 3.10 or higher, pipx, and several system libraries that OpenCV depends on: ffmpeg, libsm6, libxext6, and libgl1-mesa-glx. On standard desktop Linux or macOS these are usually present, but they are often missing in minimal Docker containers or cloud virtual machines.

For Ubuntu or Debian systems the README provides the install command:

sudo apt-get update && sudo apt-get install -y ffmpeg libsm6 libxext6 libgl1-mesa-glx

For Docker, the same packages are added via a RUN instruction in the Dockerfile.

Object Detection and OCR in Practice

The detect and find tools use models from Ultralytics, which includes the YOLO model family. The models directory is managed by the server, and the models://list resource shows which models are available locally. Downloading a specific model, such as foduucom/web-form-ui-field-detection mentioned in the README's example, requires referencing it by name in the prompt or through the config tool.

The README gives a practical example for form field detection: "Enumerate form fields on this form.jpg with foduucom/web-form-ui-field-detection model and fill the form.md with a list of described fields". The README includes a hint to specify the model and the confidence threshold explicitly.

OCR via EasyOCR handles text extraction from images. EasyOCR supports multiple languages, and the language parameter is passed in the prompt. The tool is suited for extracting text from screenshots, photographs of documents, or any image where the text is not machine-readable in the original file format.

The README includes a hint that applies to all tools: adding "use imagesorcery" to a prompt ensures the AI assistant selects the correct tool when multiple MCP servers are available and the task could be handled by more than one of them.

Privacy Model and Processing Architecture

Every tool in ImageSorcery MCP processes images locally. The README explicitly states this in the feature list: "Do all of this locally, without sending your images to any servers". This contrasts with cloud-based image analysis APIs, where images are uploaded to a remote service for processing.

The local model means the only data leaving the machine is the text of the tool invocation itself, which passes through the AI assistant's normal conversation channel. The image files remain on disk and are read by the locally running MCP server process.

Model weights for Ultralytics detection models and EasyOCR are downloaded from Hugging Face during first use. The huggingface_hub package is listed as a dependency in pyproject.toml, and tqdm handles progress display during model downloads. After download, inference runs locally.

The package includes telemetry dependencies: amplitude-analytics and posthog are both listed in pyproject.toml. The README does not document what telemetry is collected or how to disable it. Users with strict data policies should review this before deploying in a sensitive environment.

Limitations and Where ImageSorcery MCP Falls Short

The package version in pyproject.toml is 0.12.0 with a development status classifier of Alpha. The Alpha classification is a meaningful signal: the API and tool behavior can change between versions without a stability guarantee.

The dependency surface is substantial. The pyproject.toml lists fastmcp, pydantic, opencv-python, imutils, Pillow, ultralytics, requests, tqdm, huggingface_hub, easyocr, toml, amplitude-analytics, posthog, and python-dotenv. Each of these has its own version requirements and potential conflicts with other Python packages in the same environment. Pipx isolation handles most of this for standalone use, but integrating into an existing Python environment requires careful dependency management.

The last push to the repository was on 2026-05-19, roughly four months before this review. The repository is not archived, but the pace of updates appears slower than during active initial development.

For complex multi-step image analysis workflows, the tool invocation needs to be explicit and specific. The README notes that the AI assistant will combine multiple tools to achieve a goal, but the quality of chaining depends on how precisely the task is described and how well the assistant maps that description to available tools.

Alternatives and Supported Clients

ImageSorcery MCP targets three documented MCP clients: Claude.app, Cline, and Cursor. The README lists each in the requirements section and provides Cline-specific installation instructions. Any MCP-compatible client that supports the standard server configuration format should work, but the documentation focuses on these three.

Comparable tools in the search data include Photopea-MCP (which provides Photoshop-style operations through Photopea's engine, a browser-based tool that processes images remotely) and YOLO MCP server (which focuses specifically on object detection). ImageSorcery MCP covers more ground than a detection-only server and differs from Photopea-MCP in that all processing is local.

For developers who already write Python scripts directly, the main value proposition is the natural language interface rather than novel functionality. The underlying libraries (OpenCV, EasyOCR, Ultralytics) are all standard tools available independently. ImageSorcery MCP is a connector layer that makes those tools accessible through MCP-compatible AI assistants without writing code.

Editorial conclusion

ImageSorcery MCP is a practical choice for developers using Claude, Cursor, or Cline who want to automate local image tasks through natural language rather than writing OpenCV scripts directly. The local processing model is its main advantage for privacy-sensitive work, since images do not leave the machine. The weak point is the system dependency list: OpenCV requires ffmpeg, libsm6, libxext6, and libgl1-mesa-glx, which adds installation friction in Docker or minimal Linux environments. Before adopting, verify that pipx can install the package cleanly on your OS and that your MCP client supports the server configuration format that imagesorcery-mcp --post-install produces.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. sunriseapps/imagesorcery-mcp on GitHub