MCP 服务器

io.github.EntityAdam/comfyui-mcp

Gives AI agents access to ComfyUI generation, workflows, installed models, recent assets, and job controls through the Model Context Protocol.

README

ComfyUI MCP Server for .NET

Give an AI agent access to ComfyUI generation, workflows, installed models, recent assets, and job controls through the Model Context Protocol.

Quickstart

1. Install .NET 10

Install the .NET 10 SDK. The SDK includes dnx, which downloads and runs the pinned NuGet package without a permanent installation.

Start ComfyUI normally and leave it available at http://localhost:8188.

2. Paste the server configuration

Add this server to your MCP configuration:

{
  "mcpServers": {
    "comfyui": {
      "command": "dnx",
      "args": [
        "ComfyUi.Mcp@0.0.1-preview5",
        "--yes",
        "--",
        "--stdio"
      ]
    }
  }
}

See MCP client configuration for a complete configuration and copy-paste alternatives, including connecting to ComfyUI at another URL.

Some clients call the top-level property servers instead of mcpServers. Restart or refresh the MCP client after saving the configuration.

3. Let your agent finish setup

Ask:

Run setup_guide() and help me configure ComfyUI MCP.

The guide discovers installed models and bundled workflow families, then helps the agent choose workflows and create reusable defaults. The server can start while ComfyUI is offline, but model discovery and generation require ComfyUI to be running.

Setup guide

For a more deliberate first setup, ask your agent to:

Run setup_guide().
Run workflow_list() and show workflows that match your installed models.
Run family_list() and family_model_list() to check model associations.
Run workflow_options_show() before the first execution.
Save any defaults you approve with workflow_config_set().
Generate a small test image and display it with image_show().

Workflow JSON defines the available parameters. Optional YAML layers provide reusable family, workflow, and model defaults. Tool-call values always have the highest precedence. See Configuration for the complete model.

Features and usage

Families

Families group compatible installed models and enable random selection.

family_list() discovers configured families.
family_model_list() shows installed members.
family_model_add() and family_model_remove() maintain exact membership.

Try asking: “Show my configured families and which installed models belong to each one.”

Workflows

Bundled workflows cover image, audio, video, editing, upscaling, interpolation, and text generation.

workflow_list() discovers workflow IDs.
workflow_options_show() resolves inputs, defaults, and allowed values.
workflow_execute() runs any parameterized workflow.
workflow_config_set() saves reusable defaults.

Try asking: “List the workflows I can run, then explain the options for the best text-to-image workflow.”

Models

model_list() reads the current checkpoint and diffusion-model inventory directly from ComfyUI. Exact model requests must use the returned filename, including its extension.

Try asking: “List my installed models and recommend compatible workflow families without inventing filenames.”

Assets

Generated outputs and their provenance are stored in a bounded SQLite registry that survives MCP server restarts.

asset_list() and asset_show() inspect recent outputs.
generation_show() follows durable event-driven generation state.
image_show() renders an image in the conversation.
asset_regenerate() reuses a submitted workflow with new parameters.
asset_publish_info() and asset_publish() copy outputs into a project.

Try asking: “Show my five most recent assets and display the newest image.”

ComfyUI

The server can upload inputs and manage ComfyUI jobs directly.

comfy_media_upload() stages image, audio, or video inputs.
comfy_queue_status_show() reports running and pending jobs.
comfy_job_show() reads local state for a prompt submitted by this server.
comfy_job_cancel() cancels queued or running work.

Try asking: “Show the ComfyUI queue and explain what is currently running.”

How-to guides

Tips and tricks

Pick a random model from a family

Ask the agent to call family_model_list() first, then use model_strategy: random with the chosen family:

{
  "prompt": "A cinematic city at dusk",
  "workflow_id": "image_krea2_turbo_t2i",
  "family": "krea2-turbo",
  "model_strategy": "random",
  "aspect_ratio": "16:9"
}

The server selects only from installed models associated with that family.

Add `readme()` to your system prompt

The readme() tool returns compact operating instructions for agents, including workflow discovery, exact model selection, media upload, and image display. Call it once and add the returned text to your agent's system prompt or persistent project instructions. read_skill() is an alias.