uncomfymcp

A Model Context Protocol (MCP) server that connects an AI assistant to a running ComfyUI instance. It fetches a workflow saved in ComfyUI, writes the prompt and seed into it, runs it, and returns the image inline in the chat — so it only works with text-to-image workflows that have a single prompt node and seed to inject into.

This is a third-party MCP server for ComfyUI. The official one, comfy-mcp, is available at https://github.com/Comfy-Org/comfy-mcp

Install

Requires Python 3.10+ and a running ComfyUI.

Linux and macOS:

git clone https://github.com/aschet/uncomfymcp.git
cd uncomfymcp
python3 -m venv .venv
.venv/bin/pip install -e .

Windows:

git clone https://github.com/aschet/uncomfymcp.git
cd uncomfymcp
python -m venv .venv
.venv\Scripts\pip install -e .

That gives you the command .venv/bin/uncomfymcp (.venv\Scripts\uncomfymcp.exe on Windows), which a client launches for you — nothing to start by hand beyond ComfyUI itself.

Related MCP server: comfyui-mcp

Connecting a Client

Any MCP client works. By default the server speaks stdio, meaning the client launches .venv/bin/uncomfymcp itself and talks to it over the process pipes. With --transport http it instead listens on http://127.0.0.1:8000/mcp for clients that connect over the network — with no authentication, so don't expose it without one in front.

Add this mcpServers entry to the client's config. Flags go in an "args" array, e.g. "args": ["--comfy-url", "http://host:8188", "--timeout", "600"] — add the first if ComfyUI isn't on localhost, and raise the timeout if your workflows are slow or an agent chains several generations.

{
  "mcpServers": {
    "uncomfy": {
      "command": "/path/to/uncomfymcp/.venv/bin/uncomfymcp"
    }
  }
}

On Windows, command is the .venv\Scripts\uncomfymcp.exe path instead.

• Claude Desktop: add it to claude_desktop_config.json.

• AnythingLLM: add it to anythingllm_mcp_servers.json (~/.config/anythingllm-desktop/storage/plugins/ on Linux), then start it from Settings → Agent Skills → MCP Servers and invoke with @agent. It renders no image at all, so add --no-inline-image to "args" — otherwise every generation wastes tens of thousands of tokens on base64 the agent can't even display.

Use

Claude Desktop

you:    which image workflows do I have?
claude: [list_workflows]
        Ready to generate with:
          Krea2
          Z-Image Turbo

        Present but not usable:
          Ideogram 4 -- needs models that are not installed:
          flux2-vae.safetensors, gemma4_e4b_it_fp8_scaled.safetensors,
          ideogram4_fp8_scaled.safetensors

you:    generate a red fox in deep snow using Krea2
claude: [image]  Krea2 · seed 12345 ·
        http://127.0.0.1:8188/view?filename=Krea2_turbo_00007_.png

The image shows inside the collapsed tool card — expand it, or use the URL.

AnythingLLM

Tool calls need the @agent prefix, and with --no-inline-image set (see above) there is no [image] block, only the line with the seed and URL:

you:    @agent generate a red fox in deep snow using Krea2
agent:  Krea2 · seed 12345 ·
        http://127.0.0.1:8188/view?filename=Krea2_turbo_00007_.png

Two tools are exposed:

Limitations

• Only the prompt and seed change — no width, height, steps or sampler; those come from the workflow. A sampler set to "fixed" in ComfyUI is not honoured.

• Node detection can pick the wrong node when a workflow has several prompt boxes; set a node's Title to MCP:prompt to override.

• Images are sent as WebP, downscaled and compressed to fit a 1 MB limit, with transparency preserved. The full-resolution original stays in ComfyUI's output folder.

Configuration

All settings are command-line flags.