artokun/comfyui-mcp

Local-first, agent-native control plane for ComfyUI — MCP server + sidebar agent that generates images, video & audio, authors and runs workflows, and edits your live graph in natural language on ANY LLM (Claude, ChatGPT, Gemini, offline Ollama, or any hosted model). 178 tools, 36 AI skills, 55 installer packs. Local, LAN, VPS, or Comfy Cloud.

506 ★83 forksTypeScriptUpdated 1d ago

What you need to know

Local-first, agent-native control plane for ComfyUI — an MCP server + Claude Code plugin + sidebar agent panel that generates images/video/audio, authors and runs workflows, manages models and custom nodes, and edits your live ComfyUI graph in natural language, working with Claude, ChatGPT, Gemini, hosted models, or a free local model via Ollama.

Install

npx -y comfyui-mcp   # in mcpServers config (e.g. ~/.claude/settings.json)
/plugin marketplace add artokun/comfyui-mcp && /plugin install comfy   # Claude Code plugin
npx -y comfyui-mcp@latest connect   # panel-orchestrator for the ComfyUI Agent Panel (sidebar agent)

Usage

  • With ComfyUI running, ask Claude to generate an image — it finds (or downloads) a checkpoint, builds a workflow, executes it, and returns the image.
  • Slash commands: /comfy:gen <prompt>, /comfy:debug [prompt_id], /comfy:batch (parameter sweeps over cfg/steps/sampler/seed/…), /comfy:recipe <name> (portrait, hires-fix, style-transfer, product-shot), /comfy:convert, /comfy:install <pack>, /comfy:gallery, /comfy:compare, /comfy:viz, /comfy:node-skill.
  • Point at any ComfyUI: `npx -y comfyui-mcp@latest --comfyui-url http://192.168.1.50:8188`; remote mode via --force-remote; cloud mode via COMFYUI_API_KEY (Comfy Cloud).
  • Expose as a hosted connector with one command: `npx -y comfyui-mcp@latest --tunnel` (streamable-HTTP + cloudflared tunnel + auth token).

Key features

  • 37 MCP tools + 37 AI skills (model-family guides for Flux, WAN, LTX 2.3, Qwen, Z-Image, Ideogram 4, ERNIE, ANIMA), 55 installer packs, 11 slash commands, 3-4 autonomous agents, 3 hooks.
  • Graph-level control: create_workflow (create/modify/validate), visualize_workflow (Mermaid rendering + DSL round-trip), enqueue_workflow, queue management, get_history diagnose.
  • Model management: download_model (URL/HuggingFace/CivitAI search), list_local_models, clear_vram; custom-node install via registry (comfy-cli integration).
  • Generation tracker (SQLite, `npm run generations:stats`) records settings combos with reuse counts; community model-settings presets for 10+ model families.
  • Transports: stdio by default, streamable-HTTP with --http; compact 3-meta-tool mode for small local models; first-class setup for Hermes, OpenClaw, Copilot CLI (`npx -y comfyui-mcp setup <agent>`).

Best for

AI-driven ComfyUI workflow authoring and image/video generation for Claude/ChatGPT/Gemini users — including fully offline use with a free local LLM via Ollama — on your own (local, LAN, VPS, or Comfy Cloud) install.

Caveats

  • Requires Node.js >= 22 and a running ComfyUI (or a remote/cloud target); tool calling is a hard requirement for the model (no tool calling = doesn't work); thinking and vision strongly recommended.
  • Cloud mode (COMFYUI_API_KEY) routes via cloud.comfy.org; WebSocket and local-only tools throw CLOUD_UNSUPPORTED.
  • Optional API tokens: CIVITAI_API_TOKEN, HUGGINGFACE_TOKEN, GITHUB_TOKEN; paid API credits are guarded (check_runtime classifies workflows as local vs api).
  • Library/package is not affiliated with Comfy-Org; Comfy-Org ships its own Comfy Cloud MCP / In-App Agent for zero-setup cloud use.
Platforms: macOS · Linux · WindowsClients: Claude Code · Claude Desktop · Cursor · Hermes · OpenClaw · Copilot CLI · Codex (ChatGPT subscription) · Gemini (planned)
Documentation ↗

Reviewed 2026-08-07

Topics

agent-skillsai-agentclaude-codeclaude-plugincomfyuicomfyui-extensioncomfyui-mcpcomfyui-mcp-serverfluximage-generationlocal-firstmcpmcp-servermodel-context-protocolofflineollamaself-hostedstable-diffusionvideo-generationwan
View on GitHub ↗
Stars
506★
Forks
83
Language
TypeScript
License
MIT
Created
2026-02-15
Last push
2026-08-06