Pixelle MCP: Brush that turns ComfyUI workflows into AI agents with one clickPixelle MCP：把 ComfyUI 工作流一键变成 AI 智能体的画笔

ComfyUI's node workflow is powerful, but it is designed for “human drag nodes”. When agents (Claude Code, Cursor, various MCP clients) want to use these capabilities, the traditional approach is to encapsulate APIs, processing parameters, and queues themselves, and the threshold is not low.

The idea of the Pixelle MCP is "AIGC is completely intelligent": read the workflow JSON in ComfyUI, automatically analyze which parameters are exposed (prompt, width, height, number of steps), and register the entire workflow as a standard MCP tool. The agent calls it like a normal function.

The official slogan is "Build your own AIGC solution in one minute", and workflow-level conversion means that you can reuse the plotted recipe in ComfyUI, and the agent can be reused as it is.

Vincent diagram, Tucson diagram, Vincent video, Tucson video, speech synthesis, speech recognition - official examples cover TTS (CosyVoice), digital human video (Wan2), Flux graphics, Qwen-VL understanding and many other workflows.

Automatic derivation of parameters: The input parameters of the tool are directly derived from the workflow node. The agent can see the type, default value and description of each parameter in the MCP Inspector (see the official screenshot: t2i_by_flux_model tool with prompt, width, height parameters).

OpenAI-compatible interface and WebUI dual channel: It can not only use the MCP protocol for the agent, but also use HTTP as an ordinary AIGC service; it comes with the Manager management model and workflow.

The environment requires Python 3.11 + with a running ComfyUI. Download the source code package (5.82 MB) to extract, or git clone official repository; configure the ComfyUI address in .env after installing dependencies.

Once started, add the SSE endpoint with the MCP client (default http://localhost: 8002/sse) and the client will automatically list all registered workflow tools.

Verification: In the MCP Inspector, you can see that tools such as t2i_by_flux_model are successful; you can also try them directly with the official WebUI. The project has README_CN.md Chinese documents, and basically does not step on the pit according to the path.

To run Pixelle MCP, you need to have deployed ComfyUI and configured the corresponding model (such as Flux). Please estimate the memory and disk usage in advance. This package is the main branch source code, licensed by mit, and the functions are subject to the latest instructions of the official warehouse.