ComfyUI MCP

ComfyUI MCP is an open-source MCP server and Claude Code plugin built by @artokun, a regular contributor to the Comfy-Org frontend and Python core. It connects AI assistants — Claude, ChatGPT, Gemini, Ollama, or any OpenAI-compatible model — to a running ComfyUI instance, whether local, remote, or on Comfy Cloud, and lets them operate the full pipeline in natural language. The project is published on npm as comfyui-mcp and is available under the MIT license.

What It Is

ComfyUI MCP is described by its author as "not a thin prompt-to-image bridge — a full agent that operates ComfyUI for you." Where most ComfyUI MCP servers forward a prompt and return an image, this project exposes 38 MCP tools that cover the entire ComfyUI surface: workflow authoring and execution, model search and download, custom node management, VRAM control, process control, diagnostics, and a compact DSL for round-tripping workflow graphs. A Claude Code plugin layer adds 42 AI skills, 11 slash commands, 4 autonomous agents, and 3 hooks on top of those tools.

Tool Surface and Capabilities

The 38 MCP tools are organized into functional groups:

• Generation — generate_image with actions for text-to-image, ControlNet, IP-Adapter, video (LTX-2.3), 3D, upscale, background removal, and audio (ACE Step 1.5, Stable Audio 3)
• Workflow authoring — create_workflow (templates: txt2img, img2img, upscale, inpaint, controlnet, ip_adapter, audio), visualize_workflow (Mermaid render, DSL round-trip), get_workflow (strip/slice to de-virtualize Get/Set buses and rgthree toggle monoliths)
• Model management — download_model (HuggingFace, CivitAI), list_local_models (list, remove, path management)
• Custom nodes — install, update, snapshot, bisect via ComfyUI-Manager; search_custom_nodes against the Comfy Registry
• Process and memory — restart_comfyui, clear_vram, get_system_stats, get_history with full Python tracebacks and a diagnose action
• Generation tracker — every enqueue_workflow call logs to a local SQLite database; get_history action:"suggest" surfaces proven sampler/CFG/steps combos from local history

The Claude Code plugin adds slash commands (/comfy:gen, /comfy:debug, /comfy:batch, /comfy:recipe, /comfy:viz, and more), four autonomous agents (explorer, debugger, optimizer, researcher), and hooks including a VRAM watchdog and a job-completion notifier.

Deployment Model

The server targets three deployment shapes from a single config:

• Local — auto-detects a ComfyUI install and port on macOS, Linux, or Windows; no flags needed
• Remote — --comfyui-url points at a RunPod pod, VPS, LAN host, or reverse-proxied instance; --tunnel opens a Cloudflare quick tunnel and prints a ready-to-paste HTTPS URL for Claude Desktop's Custom Connectors
• Comfy Cloud — set COMFYUI_API_KEY to route requests through cloud.comfy.org

A connect subcommand runs the panel orchestrator locally and drives a remote ComfyUI from the user's own Claude or ChatGPT subscription, with no agent software installed on the remote box.

Multi-LLM and Local-Model Support

The panel orchestrator uses a provider-neutral AgentBackend port with adapters for Claude (via @anthropic-ai/claude-agent-sdk), ChatGPT/Codex (via @openai/codex), Gemini CLI, Grok, Kimi, GLM, and any Ollama or OpenAI-compatible endpoint. The author publishes a QLoRA fine-tuned Gemma 4 model (artokun/gemma4-comfyui-mcp) trained on 1,055 server-verified trajectories for fully offline use. A compact tool mode (--compact) registers three meta-tools instead of the full set, reducing context pressure for smaller local models.

Update: v0.52.167

The latest release is v0.52.167, published on 2026-09-01. The GitHub repository shows active development with the last push on 2026-09-02. The project has accumulated 711 stars and 109 forks since its creation in February 2026. Recent additions documented in the README include the ComfyUI Agent Panel (available on ComfyUI-Manager and the Comfy Registry), 56 installer packs covering model families such as Flux, WAN, LTX-2.3, MiniMax H3, Qwen, Z-Image, Ideogram 4, ERNIE, and ANIMA, and the kitchen tool for GPU-specific performance assessment. The --tunnel flag for remote/hosted connector mode and the comfy_cli integration (requiring comfy-cli ≥ 1.11.1) are also recent additions.