MCP 服务器

comfyops-mcp

Wraps ComfyUI to generate images, video, and upscales from text prompts via MCP tools, providing a curated workflow interface without the node editor.

README

comfyops-mcp — Local Generative AI Engine

Wraps ComfyUI so you can generate images, video, and upscales from a prompt — without ever touching the node editor. Pick a curated workflow, type what you want, get a PNG or MP4.

Comfy for the Perplexed

What's ComfyUI? It's the engine room of local AI image generation. You connect boxes labelled "KSampler", "CLIP Text Encode", and "VAE Decode" with wires, like a Moog synthesizer made by someone who hates knobs. It's incredibly powerful and completely impenetrable.

What's this? comfyops-mcp is a butler for ComfyUI. You tell it what you want — "a cat surfing on a pizza slice" — and it picks the right boxes, wires them up, and hands you the result. No node editor, no patch cables, no "what's a CFG scale?"

What can it do?

Turn text into images (FLUX.2, SDXL)
Turn text or photos into short videos (Wan 2.2)
Upscale old low-res images to 4K (ESRGAN)
Inpaint: replace parts of an image ("make that car red")
Browse and catalog workflows from the community
Remember every generation with its seed so you can reproduce it later

Who is it for?

You, who just wants pictures — prompt in, PNG out
You, who has an RTX 4090 gathering dust — put it to work
You, who already uses ComfyUI — now your AI assistant can drive it
You, who saw a cool AI image online and wants to make one

If you already know ComfyUI: comfyops registers curated workflows from workflows/ as MCP tools. You can add your own with comfy_workflows/register. The workflows dir is yours — export from the node editor, drop in, done.

Preview

ComfyUI Node Editor (what we wrap)	comfyops Dashboard (what you see)

Screenshots pending — run just screenshots once the dashboard is live.

What You Can Do

"Generate a 1024x1024 image of a cat surfing on a pizza slice"
"Upscale my archive of 2024 SD generations to 4K"
"Turn this photo into a cinematic video" (Wan 2.2 i2v)
"Make 10 variations of this character concept"

Quick Install

git clone https://github.com/sandraschi/comfyops-mcp
cd comfyops-mcp
uv sync
.\start.ps1

See INSTALL.md for prerequisites and Claude Desktop config.

Features

Prompt-to-image: FLUX.2 klein, SDXL, Z-Image Turbo — curated workflows with sensible defaults
Prompt-to-video: Wan 2.2 (quality) and LTX-Video (speed)
Upscale & restore: ESRGAN and SUPIR for archive restoration
Inpaint & edit: Region-based editing via FLUX.2
VRAM guard: Checks free GPU memory before queueing — no OOM crashes
Seed control: Same seed + same workflow = identical output, every time
Generation library: SQLite history with search — browse past prompts and seeds
Agentic workflow: Multi-step generation planning via MCP sampling (SEP-1577)
Prefab cards: Rich in-chat status and generation cards

How It Runs

Mode	ComfyUI required?	When
MCP tools (stdio/HTTP)	Yes	Generation via Claude Desktop, Cursor, opencode
Dashboard (Vite)	No	Browse workflows, models, library — generation needs ComfyUI
ComfyUI sidecar	Optional	`start.ps1` can launch ComfyUI automatically if `COMFYOPS_COMFYUI_DIR` is set

Workflow Depot

comfyops ships with 6 curated workflows, but the community has thousands. Use comfy_workflows/discover to pull from community sources, or comfy_workflows/register to add your own exports from the ComfyUI node editor. The webapp has a full recipe browser with tags, metadata, and search.

Tools

Tool	Ops
`comfy_generate`	image, video, upscale, inpaint, edit
`comfy_workflows`	list, get, validate, register, search, discover
`comfy_models`	list_installed, check_vram, health
`comfy_library`	recent, search, record
`comfy_agentic_assist`	Multi-step via MCP sampling