These dependencies are typically included with OpenCV installation and don't require separate installation. But they might be missing in some virtual environments like Docker.

For Ubuntu/Debian systems:

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

For Docker containers: Add this line to your Dockerfile:

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

0. Install pipx (if not already installed):
   bash

   Copy

   # On macOS with Homebrew:
   brew install pipx
   
   # On Ubuntu/Debian:
   sudo apt update && sudo apt install pipx
   
   # On other systems with pip:
   pip install --user pipx
   pipx ensurepath
   

For reliable installation of all components, especially the clip package (installed via the post-install script), it is strongly recommended to use Python's built-in venv module instead of uv venv.

1. Create and activate a virtual environment:

   bash

   Copy

   python -m venv imagesorcery-mcp
   source imagesorcery-mcp/bin/activate  # For Linux/macOS
   # source imagesorcery-mcp\Scripts\activate    # For Windows
   

2. Install the package into the activated virtual environment: You can use pip or uv pip.

   Terminal

   Copy

   pip install imagesorcery-mcp
   # OR, if you prefer using uv for installation into the venv:
   # uv pip install imagesorcery-mcp
   

3. Run the post-installation script: This step is crucial. It downloads the required models and attempts to install the clip Python package from GitHub into the active virtual environment.

   bash

   Copy

   imagesorcery-mcp --post-install
   

Note: When using this method, you'll need to provide the full path to the executable in your MCP client configuration (e.g., /full/path/to/venv/bin/imagesorcery-mcp).

• Creates a config.toml configuration file in the current directory, allowing users to customize default tool parameters.
• Creates a models directory (usually within the site-packages directory of your virtual environment, or a user-specific location if installed globally) to store pre-trained models.
• Generates an initial models/model_descriptions.json file there.
• Downloads default YOLO models (yoloe-11l-seg-pf.pt, yoloe-11s-seg-pf.pt, yoloe-11l-seg.pt, yoloe-11s-seg.pt) required by the detect tool into this models directory.
• Attempts to install the clip Python package from Ultralytics' GitHub repository directly into the active Python environment. This is required for text prompt functionality in the find tool.
• Downloads the CLIP model file required by the find tool into the models directory.

You can run this process anytime to restore the default models and attempt clip installation.

• Using uv venv to create virtual environments: Based on testing, virtual environments created with uv venv may not include pip in a way that allows the imagesorcery-mcp --post-install script to automatically install the clip package from GitHub (it might result in a "No module named pip" error during the clip installation step). If you choose to use uv venv:

  1. Create and activate your uv venv.
  2. Install imagesorcery-mcp: uv pip install imagesorcery-mcp.
  3. Manually install the clip package into your active uv venv:
     bash

     Copy

     uv pip install git+https://github.com/ultralytics/CLIP.git
     

  4. Run imagesorcery-mcp --post-install. This will download models but may fail to install the clip Python package. For a smoother automated clip installation via the post-install script, using python -m venv (as described in step 1 above) is the recommended method for creating the virtual environment.

• Using uvx imagesorcery-mcp --post-install: Running the post-installation script directly with uvx (e.g., uvx imagesorcery-mcp --post-install) will likely fail to install the clip Python package. This is because the temporary environment created by uvx typically does not have pip available in a way the script can use. Models will be downloaded, but the clip package won't be installed by this command. If you intend to use uvx to run the main imagesorcery-mcp server and require clip functionality, you'll need to ensure the clip package is installed in an accessible Python environment that uvx can find, or consider installing imagesorcery-mcp into a persistent environment created with python -m venv.

For pipx installation (recommended):

"mcpServers": {
    "imagesorcery-mcp": {
      "command": "imagesorcery-mcp.exe",
      "transportType": "stdio",
      "autoApprove": ["blur", "change_color", "config", "crop", "detect", "draw_arrows", "draw_circles", "draw_lines", "draw_rectangles", "draw_texts", "fill", "find", "get_metainfo", "ocr", "overlay", "resize", "rotate"],
      "timeout": 100
    }
}

For manual venv installation:

"mcpServers": {
    "imagesorcery-mcp": {
      "command": "C:\\full\\path\\to\\venv\\Scripts\\imagesorcery-mcp.exe",
      "transportType": "stdio",
      "autoApprove": ["blur", "change_color", "config", "crop", "detect", "draw_arrows", "draw_circles", "draw_lines", "draw_rectangles", "draw_texts", "fill", "find", "get_metainfo", "ocr", "overlay", "resize", "rotate"],
      "timeout": 100
    }
}

When downloading models, the script automatically updates the models/model_descriptions.json file:

• For Ultralytics models: Descriptions are predefined in src/imagesorcery_mcp/scripts/create_model_descriptions.py and include detailed information about each model's purpose, size, and characteristics.

• For Hugging Face models: Descriptions are automatically extracted from the model card on Hugging Face Hub. The script attempts to use the model name from the model index or the first line of the description.

After downloading models, it's recommended to check the descriptions in models/model_descriptions.json and adjust them if needed to provide more accurate or detailed information about the models' capabilities and use cases.

1. STDIO Mode (Default) - This is the standard mode for local MCP clients:

   bash

   Copy

   imagesorcery-mcp
   

2. Streamable HTTP Mode - For web-based deployments:

   bash

   Copy

   imagesorcery-mcp --transport=streamable-http
   

   With custom host, port, and path:

   bash

   Copy

   imagesorcery-mcp --transport=streamable-http --host=0.0.0.0 --port=4200 --path=/custom-path
   

Available transport options:

• --transport: Choose between "stdio" (default), "streamable-http", or "sse"
• --host: Specify host for HTTP-based transports (default: 127.0.0.1)
• --port: Specify port for HTTP-based transports (default: 8000)
• --path: Specify endpoint path for HTTP-based transports (default: /mcp)

Directory Structure

This repository is organized as follows:

.
├── .gitignore                 # Specifies intentionally untracked files that Git should ignore.
├── pyproject.toml             # Configuration file for Python projects, including build system, dependencies, and tool settings.
├── pytest.ini                 # Configuration file for the pytest testing framework.
├── README.md                  # The main documentation file for the project.
├── setup.sh                   # A shell script for quick setup (legacy, for reference or local use).
├── models/                    # This directory stores pre-trained models used by tools like `detect` and `find`. It is typically ignored by Git due to the large file sizes.
│   ├── model_descriptions.json  # Contains descriptions of the available models.
│   ├── settings.json            # Contains settings related to model management and training runs.
│   └── *.pt                     # Pre-trained model.
├── src/                       # Contains the source code for the 🪄 ImageSorcery MCP server.
│   └── imagesorcery_mcp/       # The main package directory for the server.
│       ├── README.md            # High-level overview of the core architecture (server and middleware).
│       ├── __init__.py          # Makes `imagesorcery_mcp` a Python package.
│       ├── __main__.py          # Entry point for running the package as a script.
│       ├── logging_config.py    # Configures the logging for the server.
│       ├── server.py            # The main server file, responsible for initializing FastMCP and registering tools.
│       ├── middleware.py        # Custom middleware for improved validation error handling.
│       ├── logs/                # Directory for storing server logs.
│       ├── scripts/             # Contains utility scripts for model management.
│       │   ├── README.md        # Documentation for the scripts.
│       │   ├── __init__.py      # Makes `scripts` a Python package.
│       │   ├── create_model_descriptions.py # Script to generate model descriptions.
│       │   ├── download_clip.py # Script to download CLIP models.
│       │   ├── post_install.py  # Script to run post-installation tasks.
│       │   └── download_models.py # Script to download other models (e.g., YOLO).
│       ├── tools/               # Contains the implementation of individual MCP tools.
│       │   ├── README.md        # Documentation for the tools.
│       │   ├── __init__.py      # Makes `tools` a Python package.
│       │   └── *.py           # Implements the tool.
│       ├── prompts/             # Contains the implementation of individual MCP prompts.
│       │   ├── README.md        # Documentation for the prompts.
│       │   ├── __init__.py      # Makes `prompts` a Python package.
│       │   └── *.py           # Implements the prompt.
│       └── resources/           # Contains the implementation of individual MCP resources.
│           ├── README.md        # Documentation for the resources.
│           ├── __init__.py      # Makes `resources` a Python package.
│           └── *.py           # Implements the resource.
└── tests/                     # Contains test files for the project.
    ├── test_server.py         # Tests for the main server functionality.
    ├── data/                  # Contains test data, likely image files used in tests.
    ├── tools/                 # Contains tests for individual tools.
    ├── prompts/               # Contains tests for individual prompts.
    └── resources/             # Contains tests for individual resources.

Development Setup

1. Clone the repository:

git clone https://github.com/sunriseapps/imagesorcery-mcp.git # Or your fork
cd imagesorcery-mcp

2. (Recommended) Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate # For Linux/macOS
# venv\Scripts\activate    # For Windows

3. Install the package in editable mode along with development dependencies:

pip install -e ".[dev]"

This will install imagesorcery-mcp and all dependencies from [project.dependencies] and [project.optional-dependencies].dev (including build and twine).

Rules

These rules apply to all contributors: humans and AI.

0. Read all the README.md files in the project. Understand the project structure and purpose. Understand the guidelines for contributing. Think through how it relates to your task, and how to make changes accordingly.

1. Read pyproject.toml. Pay attention to sections: [tool.ruff], [tool.ruff.lint], [project.optional-dependencies] and [project]dependencies. Strictly follow code style defined in pyproject.toml. Stick to the stack defined in pyproject.toml dependencies and do not add any new dependencies without a good reason.

2. Write your code in new and existing files. If new dependencies are needed, update pyproject.toml and install them via pip install -e . or pip install -e ".[dev]". Do not install them directly via pip install. Check out existing source codes for examples (e.g. src/imagesorcery_mcp/server.py, src/imagesorcery_mcp/tools/crop.py). Stick to the code style, naming conventions, input and output data formats, code structure, architecture, etc. of the existing code.

3. Update related README.md files with your changes. Stick to the format and structure of the existing README.md files.

4. Write tests for your code. Check out existing tests for examples (e.g. tests/test_server.py, tests/tools/test_crop.py). Stick to the code style, naming conventions, input and output data formats, code structure, architecture, etc. of the existing tests.

5. Run tests and linter to ensure everything works:

pytest
ruff check .

In case of failures - fix the code and tests. It is strictly required to have all new code to comply with the linter rules and pass all tests.

Coding hints

• Use type hints where appropriate
• Use pydantic for data validation and serialization