Pixelle-MCP is an open-source omnimodal AIGC solution based on the MCP protocol, built on top of the ComfyUI ecosystem, with the goal of seamlessly converting ComfyUI workflows into MCP tools to enable zero-code development. The project supports both local self-hosted ComfyUI and RunningHub cloud ComfyUI execution modes.
[Purpose] This framework is aimed at users who need to invoke multimodal generation capabilities in an agent-based way, and supports TISV omnimodality, that is, the conversion and generation of text, images, sound/speech, and video. It can be integrated with any MCP client and comes with a web interaction interface.
[Main Capabilities]
- Dual execution modes: local ComfyUI self-hosted environment, or RunningHub cloud ComfyUI service.
- Zero-code workflow-to-tool: by defining parameters and outputs through special DSL syntax in ComfyUI node titles, workflows can be converted into MCP tools.
- MCP server: based on the MCP protocol, it can be connected to by any MCP client such as Cursor and Claude Desktop.
- Web interface: based on the Chainlit framework, it supports multimodal interaction and can integrate more MCP servers.
- Multi-LLM support: integrates models such as OpenAI, Ollama, Gemini, DeepSeek, Claude, and Qwen through the LiteLLM framework.
- One-click deployment: supports methods such as PyPI installation, CLI commands, and Docker.
- Unified architecture: integrates the MCP endpoint, web interface, file service, and workflow engine into the same application.
[Usage] Three startup paths are provided:
- One-click experience: zero-configuration startup, after which it automatically enters the configuration wizard, guiding the selection of the execution engine (ComfyUI/RunningHub) and LLM configuration.
- Local development deployment: supports custom workflows and secondary development; the workflow must first be tested in ComfyUI to ensure it runs properly.
- Docker deployment: suitable for production environments and containerized deployment. After startup, access the web interface http://localhost:9004，默认用户名与密码均为 dev, can be modified after startup; the MCP endpoint is located at http://localhost:9004/pixelle/mcp，供 Cursor, Claude Desktop, etc. Default port is 9004, customizable via the environment variable PORT=your_port. On first launch, the configuration status is detected; you need to select an execution engine and configure at least one LLM provider, and the system will automatically create the workflow directory structure.
[Adding Custom MCP Tools] One workflow corresponds to one MCP tool, and two ways of adding are supported: use a workflow file in API format exported from local ComfyUI, or directly use a RunningHub workflow ID. Taking image Gaussian blur as an example, build the workflow in ComfyUI, set the LoadImage node title to $image.image!:Input image URL, export it as an API format file (such as i_blur.json), enter it on the web page and the LLM will automatically convert it into an MCP tool; after refreshing the page, send any image to process it.
[Workflow Customization Specification] Double-click a node title on the ComfyUI canvas to edit it, and use DSL syntax to define parameters: param_name is the MCP tool function parameter name; ~ indicates that the URL parameter is uploaded for processing and a relative path is returned; field_name is the corresponding input field of the node; ! indicates that the parameter is required; description is the parameter description. The system automatically infers the type based on the node field's current value: int (such as 512, 1024), float (such as 1.5, 3.14), bool (such as true, false), str (default). For output, the system automatically detects nodes such as SaveImage, SaveVideo, SaveAudio, VHS_SaveVideo, VHS_SaveAudio, and you can also use $output.var_name to mark output in any node title. Optionally add a text node titled MCP to provide the tool description, with the description content filled in its value field. Notes: optional parameters must have default values set in the node; fields already connected to other nodes will not be parsed as parameters; the exported file name will be used as the tool name, so it is recommended to use a meaningful English name; it must be exported in API format rather than UI format.
[Applicable Scenarios and Limitations] RunningHub cloud mode requires no local GPU or ComfyUI environment; register to obtain an API key and you can use it, with stable performance and automatic scaling; local mode gives full control over executionenvironment and model versions, local data processing, support for custom models and nodes not available in the cloud, offline operation, and no cloud service fees for high-frequency use. When using it, note that the workflow must first be tested successfully in ComfyUI, otherwise execution will fail. The project is released under the MIT license.
