ComfyUI MCP is an MCP server and Claude Code plugin through which an LLM agent operates a ComfyUI instance: it generates images, video and audio, authors and edits workflows node by node, runs and iterates on them, manages models and custom nodes, and edits the live graph in natural language. The project positions itself as a control plane rather than a prompt-to-image bridge. A self-hosted ComfyUI is the primary target, and the same configuration reaches remote installs (RunPod, a VPS, a LAN host behind a reverse proxy) and Comfy Cloud.
The server talks to ComfyUI over its REST API and WebSocket interface, and to external services for the rest: Hugging Face and CivitAI for model search and download, the ComfyUI Registry for custom nodes, and GitHub for skill generation. It speaks stdio by default and can serve streamable HTTP, including a tunneled, token-authenticated endpoint for remote clients such as Claude Desktop. One of three deployment modes (local, remote, cloud) is selected from the environment, and a compact mode registers three meta-tools instead of the full set for small local models.
The model is the user's choice: Claude or ChatGPT on a subscription, Gemini on a Google login, a local model through Ollama, or a hosted model over one API key (DeepSeek, GLM, Kimi and others via OpenRouter). Tool calling is required, and thinking and vision are recommended. A Gemma 4 family fine-tuned on server-verified tool trajectories is published for Ollama. Beyond Claude Code, a setup command writes the server entry into the Hermes, OpenClaw and Copilot CLI configs. A separate ComfyUI Agent Panel, installed from ComfyUI-Manager, embeds the same agent in the ComfyUI sidebar and asks before spending paid API credits.
The project is in limited maintenance: releases focus on verified critical fixes, and there is no broad feature roadmap.
Features
- MCP tools: 38 action-parameterized tools covering generation, assets, defaults, workflow execution, visualization, composition, the workflow library, images, models, VRAM, registry search, diagnostics and process control
- Generation: text-to-image with ControlNet and IP-Adapter conditioning, short video, 3D models through partner API nodes, audio (ACE Step 1.5, Stable Audio 3), ESRGAN upscaling and background removal
- Workflow authoring: create from templates (txt2img, img2img, upscale, inpaint, controlnet, ip_adapter), modify node by node, validate as a dry run, convert between UI and API formats, render as a Mermaid diagram and round-trip a compact DSL
- Model management: search Hugging Face and CivitAI, download into the correct ComfyUI subdirectory, list installed models by type, resolve missing files, and dedupe downloads through a content-addressed cache
- Custom nodes: search the ComfyUI Registry, install node packs, generate a Claude skill for a pack from its Registry ID or GitHub URL, and drive the official comfy-cli when it is present
- Diagnostics: server logs, execution history with tracebacks, a one-call failure diagnosis that names the failed node and the missing models, VRAM clearing and ComfyUI stop, start and restart
- Generation tracker: every run is logged to a local SQLite database keyed by model and LoRA content hash, so the agent can suggest proven sampler, scheduler, steps and CFG settings
- Claude Code plugin: 11 slash commands, 42 skills (model-family guides, a model registry, CivitAI pairing, node authoring, troubleshooting, a compatibility matrix), 4 agents (explorer, debugger, optimizer, researcher) and 3 hooks (VRAM watchdog, save warning, job completion)
- Installer packs: one-command ComfyUI setups bundling custom nodes, model URLs and a workflow for ANIMA, Ideogram 4, LTX-2.3, ERNIE, WAN, Qwen and Z-Image, with CI validating every model link
- Deployment modes: local, remote (a non-loopback URL, or a reverse proxy with its own auth header) or Comfy Cloud, auto-selected from the environment
- Transports: stdio by default, streamable HTTP, and a one-flag tunnel that prints an authenticated URL for Claude Desktop connectors and other remote clients
- Provider choice: Claude, ChatGPT, Gemini, Ollama local models or any OpenAI-compatible endpoint, with a 3-meta-tool compact mode for small local models
