ModelScope Image MCP Server

English | 中文

An MCP (Model Context Protocol) server for generating images via the ModelScope image generation API. This server provides seamless integration with AI assistants, enabling them to create images through natural language prompts with robust async processing and local file management.

IMPORTANT: Earlier drafts of this README mentioned features like returning base64 data, negative prompts, and additional parameters. The current released code (see src/modelscope_image_mcp/server.py) implements a focused minimal feature set: one tool generate_image that submits an async task and saves the resulting image locally. Planned / upcoming features are listed in the roadmap below.

Current Features

• Asynchronous image generation using ModelScope async task API
• Periodic task status polling (every 5 seconds, up to 2 minutes)
• Saves the first generated image to a local file
• Returns task status and image URL to the MCP client
• Robust error handling + timeout messaging
• Simple one-command start with uvx

Environment Variable

The server reads your credential from:

MODELSCOPE_SDK_TOKEN

If it is missing, the server will raise an error. Obtain a token from: https://modelscope.cn/my/myaccesstoken

Set on Windows (cmd):

set MODELSCOPE_SDK_TOKEN=your_token_here

PowerShell:

$env:MODELSCOPE_SDK_TOKEN="your_token_here"

Unix/macOS bash/zsh:

export MODELSCOPE_SDK_TOKEN=your_token_here

Installation & MCP Client Configuration

You can register the server directly in an MCP-compatible client (e.g. Claude Desktop) without a prior manual install thanks to uvx.

Option 1: PyPI (Recommended once published)

{
  "mcpServers": {
    "modelscope-image": {
      "command": "uvx",
      "args": ["modelscope-image-mcp"],
      "env": {
        "MODELSCOPE_SDK_TOKEN": "your_token_here"
      }
    }
  }
}

Option 2: Direct from GitHub

{
  "mcpServers": {
    "modelscope-image": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/zym9863/modelscope-image-mcp.git",
        "modelscope-image-mcp"
      ],
      "env": {
        "MODELSCOPE_SDK_TOKEN": "your_token_here"
      }
    }
  }
}

Option 3: Local Development Checkout

git clone https://github.com/zym9863/modelscope-image-mcp.git
cd modelscope-image-mcp
uv sync

Then configure MCP client entry using:

{
  "mcpServers": {
    "modelscope-image": {
      "command": "uvx",
      "args": ["--from", ".", "modelscope-image-mcp"],
      "env": { "MODELSCOPE_SDK_TOKEN": "your_token_here" }
    }
  }
}

Quick Local Smoke Test

# Run directly (local checkout)
uvx --from . modelscope-image-mcp

When running successfully you should see log lines showing task submission and polling.

## Usage Examples

### Basic Image Generation

```jsonc
{
  "name": "generate_image",
  "arguments": {
    "prompt": "A serene mountain landscape at sunset"
  }
}

Advanced Configuration

{
  "name": "generate_image",
  "arguments": {
    "prompt": "A futuristic city with flying cars, cyberpunk style",
    "model": "Qwen/Qwen-Image",
    "size": "1024x1024",
    "output_filename": "cyberpunk_city.png",
    "output_dir": "./generated_images"
  }
}

Creative Prompts

• Art Style: "in the style of Van Gogh", "watercolor painting", "digital art"
• Composition: "close-up portrait", "wide-angle landscape", "bird's eye view"
• Lighting: "dramatic lighting", "golden hour", "studio lighting"
• Mood: "mysterious atmosphere", "vibrant colors", "minimalist design"

Best Practices

1. Be Specific: Detailed prompts produce better results than vague ones
2. Use References: Mention specific art styles, artists, or time periods
3. Experiment: Try variations of your prompt to find the best result
4. Organize Outputs: Use descriptive filenames and organized directories
5. Check Status: Monitor the async task status for long-running generations

generate_image

Creates an image from a text prompt using the ModelScope async API.

Parameters:

• prompt (string, required): The text description of the desired image
• model (string, optional, default: Qwen/Qwen-Image): Model name passed to API
• size (string, optional, default: 1024x1024): Image resolution size, Qwen-Image supports: [64x64,1664x1664]
• output_filename (string, optional, default: result_image.jpg): Local filename to save the first output image
• output_dir (string, optional, default: ./outputs): Directory path where the image will be saved

Sample invocation (conceptual JSON sent by MCP client):

{
  "name": "generate_image",
  "arguments": {
    "prompt": "A golden cat playing in a garden",
    "size": "1024x1024",
    "output_filename": "cat.jpg",
    "output_dir": "./my_images"
  }
}

Sample textual response payload (returned to the client):

图片生成成功！
提示词: A golden cat playing in a garden
模型: Qwen/Qwen-Image
保存路径: /path/to/my_images/cat.jpg
输出目录: /path/to/my_images
文件名: cat.jpg
图片URL: https://.../generated_image.jpg

Notes:

• Only the first image URL is used (if multiple are ever returned)
• If the task fails or times out you receive a descriptive message
• No base64 data is currently returned (roadmap item)

Internal Flow

1. Submit async generation request with header X-ModelScope-Async-Mode: true
2. Poll task endpoint /v1/tasks/{task_id} every 5 seconds (max 120 attempts ~= 2 minutes)
3. On SUCCEED download first image and save via Pillow (PIL)
4. Return textual metadata to MCP client
5. Provide clear error / timeout messages otherwise

Roadmap

Planned enhancements (not yet implemented in server.py):

• Optional base64 return data
• Negative prompt & guidance parameters
• Adjustable polling interval & timeout via arguments
• Multiple image outputs selection
• Streaming progress notifications

Development

# Install all (including dev) dependencies
uv sync --dev

# Run server module directly
uv run python -m modelscope_image_mcp.server

# Or via uvx using local source
uvx --from . modelscope-image-mcp

# Run with environment variable
MODELSCOPE_SDK_TOKEN=your_token_here uv run python -m modelscope_image_mcp.server

# Format code (if ruff is configured)
uv run ruff format .

# Lint code (if ruff is configured)
uv run ruff check . --fix

Project Structure

modelscope-image-mcp/
├── src/modelscope_image_mcp/
│   ├── __init__.py
│   └── server.py          # Main MCP server implementation
├── pyproject.toml         # Project configuration and dependencies
├── uv.lock               # Lock file for reproducible builds
├── README.md             # This file
└── README-zh.md         # Chinese documentation

Troubleshooting

Changelog

1.0.1

• Added size parameter support for customizable image resolution
• Improved image generation with Qwen-Image model resolution range [64x64,1664x1664]
• Enhanced documentation with size parameter usage examples

1.0.0

• Major update with improved async handling and output directory support
• Added configurable output directory parameter
• Enhanced error handling and logging
• Updated dependencies to use httpx for better async support
• Fixed notification_options bug from initial release

0.1.0

• Initial minimal implementation with async polling & local image save
• Fixed bug: notification_options previously None causing AttributeError

License

MIT License

Contributing

PRs & issues welcome. Please describe reproduction steps for any failures.

Disclaimer

This is an unofficial integration example. Use at your own risk; abide by ModelScope Terms of Service.