mcp.film

mcp.film / 3D, Worlds & VFX / ComfyUI MCP (Claude Code plugin)

ComfyUI MCP (Claude Code plugin)

Communityby Community (artokun)freeverified Oct 15, 2026

Drive ComfyUI as a live image, video, audio, and workflow workbench from your agent.

What it does

TypeScript MCP server plus Claude Code plugin for ComfyUI: generate assets, execute and author workflows, manage models and custom nodes, inspect outputs, and edit the live graph. A sidebar Agent Panel turns ComfyUI itself into an agent surface while the MCP server supplies model-family skills, installer packs, and paid-credit guardrails. Runs against local, LAN, VPS, or Comfy Cloud instances.

Connect

Claude Code
claude mcp add comfyui-artokun -- npx -y comfyui-mcp@latest
Claude Desktop (claude_desktop_config.json)
{
  "mcpServers": {
    "comfyui-artokun": {
      "command": "npx",
      "args": [
        "-y",
        "comfyui-mcp@latest"
      ]
    }
  }
}
Cursor (.cursor/mcp.json)
{
  "mcpServers": {
    "comfyui-artokun": {
      "command": "npx",
      "args": [
        "-y",
        "comfyui-mcp@latest"
      ]
    }
  }
}
Auth: local-app — Requires ComfyUI running (auto-detects ports 8188/8000, or --comfyui-url for remote)

Sample tools

generate_imageenqueue_workflowbuild_workflowvisualize_workflowview_imagesearch_modelsinstall_nodestrain_loradeploy_podlist_packs

Field notes

Re-checked 15 Oct 2026: npm comfyui-mcp 0.52.206, repo pushed 14 Oct, MIT, 825 stars (715 in Sep). The release cadence has slowed — the README now says 'limited maintenance has resumed', with releases focused on verified critical fixes and no broad feature roadmap — so pin an exact version and don't count on new model-family support landing quickly. The README contradicts itself on counts (38 tools in the body, 178 in the intro). The tool surface was consolidated, not grown: the README now claims 38 tools, 42 AI skills and 56 installer packs, against 178 tools in August, because the nine generation modes (image, audio, video, 3d, controlnet, ip_adapter, regenerate, upscale, remove_background) collapsed into one action-parameterised tool at 0.50.0. Prompts and configs that called generate_audio or generate_video will break. The vendor's own docs disagree on its name — the tool page says generate_image, the llms.txt index says generate — so probe the live tool list before scripting against it. Backends beyond Claude: ChatGPT and Gemini on subscription, Ollama fully offline, DeepSeek/GLM/Kimi via OpenRouter. COMFYUI_API_KEY switches the target to Comfy Cloud instead of a local instance. --http binds a token-authed HTTP server (COMFYUI_MCP_HTTP_TOKEN); --tunnel exposes it publicly over cloudflared, which is a decision, not a convenience. Heavier than shawnrushefsky/comfyui-mcp but far more filmmaker-useful for local video/audio workflows.

Also in 3D, Worlds & VFX

Used in

Open Lab · Run a hackable open film lab

Pairs well with