Imported from reason-machines/mcp-skills (skills/comfyui-mcp-agent/SKILL.md). Install upstream with npx skills add reason-machines/mcp-skills --skill comfyui-mcp-agent. Copyright stays with the author.

comfyui-mcp-agent

Skill by ara.so — MCP Skills collection.

Expert skill for comfyui-mcp, an agent-native control plane for ComfyUI. This MCP server exposes 108 tools for image/video/audio generation, workflow authoring, model management, custom node installation, and natural language graph editing. It ships with 32 AI skills (Flux, WAN, LTX 2.3, Qwen, Ideogram 4, Krea2), 13 installer packs, and Claude Code plugin integration.

comfyui-mcp is local-first but not local-only: one config targets local installs, LAN, VPS, or Comfy Cloud. It authors and edits graphs node-by-node, runs and iterates on workflows, manages models and custom nodes, and ships model-specific expertise (samplers, CFG, resolutions, curated URLs) so agents get it right without trial and error.

Installation

As MCP Server (Claude Desktop / Claude Code)

Add to your Claude config (~/.claude/settings.json):

{
  "mcpServers": {
    "comfyui": {
      "command": "npx",
      "args": ["-y", "comfyui-mcp"],
      "env": {
        "CIVITAI_API_TOKEN": "",
        "COMFYUI_URL": "",
        "COMFYUI_API_KEY": ""
      }
    }
  }
}

Environment variables (all optional):

• CIVITAI_API_TOKEN — for downloading gated models from Civitai
• COMFYUI_URL — override auto-detected ComfyUI endpoint (default: http://127.0.0.1:8188)
• COMFYUI_API_KEY — for Comfy Cloud mode (cloud.comfy.org)
• COMFYUI_MCP_HTTP_TOKEN — auth token for remote HTTP transport (--tunnel or --http)

As Claude Code Plugin

# In Claude Code
/plugin marketplace add artokun/comfyui-mcp
/plugin install comfy

Remote / Hosted Connector (HTTP Transport)

Run as an authenticated, publicly-reachable server:

npx -y comfyui-mcp@latest --tunnel

This forces HTTP transport, generates an auth token, opens a cloudflared tunnel, and prints a https://…/mcp URL + token for Claude Desktop Custom Connectors.

Panel Agent (Autonomous Sidebar in ComfyUI)

Install the ComfyUI Agent Panel via ComfyUI-Manager:

# Search "comfyui-agent-panel" in ComfyUI-Manager
# Or manual install:
cd custom_nodes
git clone https://github.com/artokun/comfyui-mcp-panel.git
cd comfyui-mcp-panel
npm install

Then start the orchestrator from your local machine:

npx -y comfyui-mcp@latest connect

Core MCP Tools

Workflow Execution

// Generate an image from a prompt
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "generate_image",
  arguments: {
    prompt: "a sunset over mountains, cinematic lighting",
    negative_prompt: "blurry, low quality",
    width: 1024,
    height: 1024,
    cfg: 7.0,
    steps: 20,
    seed: -1,
    checkpoint: "flux1-dev-fp8.safetensors"
  }
});

// Enqueue a custom workflow (API format)
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "enqueue_workflow",
  arguments: {
    workflow: {
      "1": {
        "class_type": "CheckpointLoaderSimple",
        "inputs": { "ckpt_name": "flux1-dev-fp8.safetensors" }
      },
      "2": {
        "class_type": "CLIPTextEncode",
        "inputs": { "text": "a photo of a cat", "clip": ["1", 1] }
      }
      // ...
    }
  }
});

// Get job status
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_job_status",
  arguments: { prompt_id: "abc-123" }
});

// Cancel running job
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "interrupt_execution",
  arguments: {}
});

Graph Editing (Natural Language)

// Add a node to the current workflow
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "add_node",
  arguments: {
    node_type: "KSampler",
    params: {
      seed: 42,
      steps: 30,
      cfg: 8.0,
      sampler_name: "euler_ancestral",
      scheduler: "normal"
    }
  }
});

// Update existing node
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "update_node",
  arguments: {
    node_id: "5",
    updates: { "cfg": 9.0, "steps": 25 }
  }
});

// Delete node
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "delete_node",
  arguments: { node_id: "12" }
});

// Connect two nodes
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "connect_nodes",
  arguments: {
    source_node: "3",
    source_output: 0,
    target_node: "7",
    target_input: "image"
  }
});

// Get current workflow state
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_workflow",
  arguments: {}
});

Model Management

// List available checkpoints
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "list_checkpoints",
  arguments: {}
});

// Download a model from URL
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "download_model",
  arguments: {
    url: "https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/flux1-dev.safetensors",
    filename: "flux1-dev.safetensors",
    type: "checkpoints"
  }
});

// Download from Civitai (requires CIVITAI_API_TOKEN)
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "download_civitai_model",
  arguments: {
    model_id: "43331",
    version_id: "47274",
    filename: "majicmixRealistic_v7.safetensors",
    type: "checkpoints"
  }
});

// List LoRAs
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "list_loras",
  arguments: {}
});

// List VAEs
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "list_vaes",
  arguments: {}
});

Custom Node Management

// List installed custom nodes
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "list_custom_nodes",
  arguments: {}
});

// Install custom node pack
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "install_custom_nodes",
  arguments: {
    repo_url: "https://github.com/ltdrdata/ComfyUI-Manager.git"
  }
});

// Update custom node
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "update_custom_nodes",
  arguments: {
    node_name: "ComfyUI-Manager"
  }
});

// Uninstall custom node
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "uninstall_custom_nodes",
  arguments: {
    node_name: "ComfyUI-Impact-Pack"
  }
});

// Get node info from object_info
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_node_info",
  arguments: {
    node_type: "KSampler"
  }
});

System Management

// Get system stats (VRAM, RAM, disk)
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_system_stats",
  arguments: {}
});

// Restart ComfyUI
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "restart_comfyui",
  arguments: {}
});

// Stop ComfyUI
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "stop_comfyui",
  arguments: {}
});

// Get queue status
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_queue",
  arguments: {}
});

// Clear queue
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "clear_queue",
  arguments: {}
});

Workflow Conversion & Visualization

// Convert UI format to API format
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "convert_workflow",
  arguments: {
    workflow: { /* UI format workflow JSON */ },
    target_format: "api"
  }
});

// Visualize workflow as Mermaid diagram
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "visualize_workflow",
  arguments: {
    workflow: { /* workflow JSON */ }
  }
});

// Compare two workflows
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "compare_workflows",
  arguments: {
    workflow_a: { /* first workflow */ },
    workflow_b: { /* second workflow */ }
  }
});

Claude Code Slash Commands

When installed as a plugin, comfyui-mcp adds these slash commands:

/comfy:gen <prompt>

Generate an image from a text description. Auto-selects checkpoint, builds workflow, returns image.

/comfy:gen a portrait of a woman in cyberpunk style, neon lights

/comfy:viz <workflow>

Visualize a workflow as a Mermaid diagram with nodes grouped by category.

/comfy:viz

/comfy:node-skill <pack>

Generate a Claude skill for a custom node pack from Registry ID or GitHub URL.

/comfy:node-skill https://github.com/ltdrdata/ComfyUI-Impact-Pack

/comfy:debug [prompt_id]

Diagnose why a workflow failed — reads history, logs, traces root cause, suggests fixes.

/comfy:debug abc-123

/comfy:batch <prompt, params>

Parameter sweep generation across cfg, sampler, steps, seed, etc.

/comfy:batch "a forest scene" cfg=[6,7,8,9] steps=[20,30]

/comfy:convert <file>

Convert between UI format and API format workflows.

/comfy:convert workflow.json

/comfy:install <pack>

Install a custom node pack — git clone, pip install, optional restart.

/comfy:install ComfyUI-Manager

/comfy:gallery [filter]

Browse generated outputs with metadata — filter by date, count, or filename.

/comfy:gallery today

/comfy:compare <a vs b>

Diff two workflows side by side — shows added/removed nodes and changed parameters.

/comfy:compare workflow_v1.json workflow_v2.json

/comfy:recipe <name> <prompt>

Multi-step recipes: portrait, hires-fix, style-transfer, product-shot.

/comfy:recipe portrait "a woman with curly hair"

Model-Specific Patterns

Flux.1 (Dev / Schnell)

// Flux requires specific settings
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "generate_image",
  arguments: {
    prompt: "cinematic photo of a sunset",
    checkpoint: "flux1-dev-fp8.safetensors",
    width: 1024,
    height: 1024,
    cfg: 1.0,  // Flux works best at CFG 1.0
    steps: 20,  // Dev: 20-30, Schnell: 4-8
    sampler_name: "euler",
    scheduler: "simple"
  }
});

WAN (Video Generation)

// WAN requires ComfyUI-WAN-Animate custom nodes
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "install_custom_nodes",
  arguments: {
    repo_url: "https://github.com/artokun/ComfyUI-WAN-Animate.git"
  }
});

// Generate video with WAN
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "enqueue_workflow",
  arguments: {
    workflow: {
      // WAN workflow nodes...
      // See packs/wan-animate.json for full example
    }
  }
});

LTX Video 2.3

// Download LTX model
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "download_model",
  arguments: {
    url: "https://huggingface.co/Lightricks/LTX-Video/resolve/main/ltx-video-2.3-fp8.safetensors",
    filename: "ltx-video-2.3-fp8.safetensors",
    type: "checkpoints"
  }
});

// Install ComfyUI-LTXVideo
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "install_custom_nodes",
  arguments: {
    repo_url: "https://github.com/Lightricks/ComfyUI-LTXVideo.git"
  }
});

Qwen Image / Image Edit

// Install Qwen nodes
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "install_custom_nodes",
  arguments: {
    repo_url: "https://github.com/Qwen2-VL/ComfyUI-Qwen2VL.git"
  }
});

// Download Qwen model
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "download_model",
  arguments: {
    url: "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/resolve/main/model.safetensors",
    filename: "qwen2-vl-7b.safetensors",
    type: "checkpoints"
  }
});

Installer Packs

The project ships 13 one-command installer packs in packs/:

• anima.json — ANIMA anime generation
• ideogram4.json — Ideogram 4 text-to-image
• ltx-2.3.json — LTX Video 2.3
• ernie.json — ERNIE (Baidu) models
• wan-animate.json — WAN video animation
• wan-longer-videos.json — WAN extended video
• wan-transparent.json — WAN transparent generation
• qwen-image.json — Qwen image generation
• qwen-image-edit.json — Qwen image editing
• z-image-turbo.json — Z-Image turbo mode
• z-image-base.json — Z-Image base
• z-image-xy-plot.json — Z-Image XY plot
• artokun-flow-wan-animate.json — WAN replace/animate flow

Each pack is a manifest of custom nodes + model URLs + workflow. Apply with:

use_mcp_tool({
  server_name: "comfyui",
  tool_name: "apply_manifest",
  arguments: {
    manifest: require("./packs/flux-dev.json")
  }
});

Or use the generated installers:

# Windows
install-windows.bat flux-dev

# Linux / RunPod
bash install-runpod.sh flux-dev

Real-Time Progress Monitoring

Instead of polling get_job_status, use the WebSocket progress monitor:

# Start in background
node monitor-progress.mjs <prompt_id> &

# Reports step progress:
# "step 5/14 (36%)"
# "Completed: output_001.png"

The monitor connects to ComfyUI's WebSocket (/ws) and emits real-time updates. On completion, it prints output filenames; on error, it reports the failing node.

Troubleshooting

Common Errors

OOM (Out of Memory)

// Check VRAM before execution
const stats = use_mcp_tool({
  server_name: "comfyui",
  tool_name: "get_system_stats"
});

if (stats.vram_free_gb < 1.0) {
  console.warn("Low VRAM — consider using fp8 checkpoint or smaller resolution");
}

Missing Custom Nodes

// Error: "Cannot find node type 'XYZ'"
// Solution: Install the custom node pack
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "install_custom_nodes",
  arguments: {
    repo_url: "https://github.com/author/ComfyUI-XYZ.git"
  }
});

// Restart ComfyUI
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "restart_comfyui"
});

Model Not Found

// List available checkpoints
const checkpoints = use_mcp_tool({
  server_name: "comfyui",
  tool_name: "list_checkpoints"
});

// Download if missing
use_mcp_tool({
  server_name: "comfyui",
  tool_name: "download_model",
  arguments: {
    url: "https://...",
    filename: "model.safetensors",
    type: "checkpoints"
  }
});

Black Images / NaN Tensors

• Flux/SD3: Use CFG 1.0-3.5 (not 7-9)
• SDXL Lightning: Use 4-8 steps (not 20-30)
• FP16 on CPU: Switch to FP32 or use --force-fp16 launch arg

dtype Mismatch

RuntimeError: expected dtype Float but got Half

Solution: Use matching precision checkpoint (fp8 vs fp16) or force cast in workflow.

Launch Flags

Pass performance flags via environment or CLI:

# Low VRAM mode
python main.py --lowvram

# Force FP16
python main.py --force-fp16

# CPU offload
python main.py --normalvram

# Disable xformers
python main.py --disable-xformers

Resources

• Documentation: https://comfyui-mcp.artokun.io/docs
• GitHub: https://github.com/artokun/comfyui-mcp
• Discord: https://discord.gg/cW9arBhzCu
• ComfyUI: https://github.com/comfyanonymous/ComfyUI
• ComfyUI Registry: https://registry.comfy.org
• Agent Panel: https://github.com/artokun/comfyui-mcp-panel

Key Principles for Agents

1. Always check VRAM before heavy workflows — use get_system_stats and warn if < 1GB free
2. Use model-specific settings — Flux at CFG 1.0, SDXL Lightning at 4-8 steps, etc.
3. Auto-install missing custom nodes — if a workflow needs a node, install it first
4. Prefer curated model URLs from skills — the model-registry skill has verified download links
5. Monitor progress via WebSocket — don't poll get_job_status in a loop
6. Validate workflows before enqueueing — use get_node_info to check inputs/outputs
7. Restart after custom node installs — many nodes require a restart to register
8. Use packs for complex setups — don't manually build multi-node workflows when a pack exists
9. Check object_info for node params — ComfyUI's schema is dynamic; always verify allowed values
10. Fail gracefully with debug commands — if a workflow errors, run /comfy:debug to diagnose