ComfyUI Cookbook

Custom workflows, custom nodes, a custom MCP server, and a custom skills suite — the ComfyUI setup I used to make Cyd & Julie and Garage Force.

47 custom workflows 22 MCP tools 30+ engine profiles 8 Claude skills 457 outputs shipped
Custom workflows Custom nodes Custom MCP server Custom skills suite
ComfyUI Power Tools

The wrapper I built around ComfyUI so Claude can drive it natively.

The MCP Server

Custom-built · 22 tools exposed to Claude

I wrote a Model Context Protocol server that sits in front of ComfyUI's HTTP API and exposes its capabilities as structured tools Claude can invoke. Any Claude instance — Claude Code, Claude Desktop, a Claude-in-Chrome tab — can queue jobs, list installed models, browse LoRAs, trigger workflows from disk, fetch outputs, and check progress without leaving the conversation. It turns ComfyUI from a node editor into an AI-addressable service.

queue_promptSubmit a workflow JSON to ComfyUI and get a job ID
get_historyRetrieve status and outputs for a finished job
list_modelsEnumerate checkpoints, LoRAs, VAEs, embeddings
list_workflowsFind a saved workflow by name or purpose
upload_inputPush reference images into ComfyUI's input folder
get_outputPull a render (image or video) back from the queue

The Skills Suite

Custom-built · 8 Claude skills that compose into end-to-end creative pipelines
/comfy-briefDiscuss the creative goal, pick the right engine, plan the graph before building
/comfy-buildText description → full ComfyUI workflow JSON with node connections wired
/comfy-coachLive node-by-node walkthrough of an existing workflow via Chrome MCP
/comfy-createAutopilot — describe what you want, get a finished render. Plans, builds, runs, iterates
/comfy-enginesManage the engine knowledge base, research new models, compare, update profiles
/comfy-huntSearch Reddit, OpenArt, CivitAI, YouTube, GitHub for workflows matching a need
/comfy-installCheck what's installed, identify missing custom nodes/models, guide setup
/comfy-runExecute a workflow via the API, monitor progress, fetch outputs for review

Custom Nodes

Authored in Python · plugged into the ComfyUI graph

When a stock ComfyUI node doesn't do what a shot needs, I write one. Custom nodes I've authored slot into workflows like any other, but embed the quirks of my specific pipelines — the Garage Force segment-and-inpaint fork, the two-pass pose enforcement path, the API wrappers for commercial video engines that didn't have official ComfyUI support yet. Because the nodes live in the same Python environment as the server, they can also call out to my MCP server or touch the VFX Dashboard database, turning ComfyUI into a bridge between creative work and production data.

47 Workflows

Production library organized by category
ImageFlux, DALL-E, Image-1, Ideogram, Stable Diffusion, t2i and i2i
VideoKling, Minimax, Runway Gen-4, Veo, Pika, Pixverse, LTX, t2v, i2v, v2v
3DHunyuan3D image-to-model and multi-view, Rodin, Tripo, GLB/OBJ output
StyleIPAdapter, pose control (two-pass), ControlNet, Luma Photon style ref
Inpaint / EditDALL-E 2 inpaint, Image-1 inpaint, SAM2 segment, Moonvalley pose transfer
UpscaleHi-res fix with ESRGAN, latent upscale variants
Engine Registry

The 30+ engine profiles I maintain, and the routing recommendations I apply when picking one for a shot.

Photo-real still→ Flux (primary), or Image-1 when a subject needs strong text fidelity
3D from single photo→ Hunyuan3D image-to-model (fast) or Rodin (higher geo quality)
3D from multiple angles→ Hunyuan3D multi-view turbo for Cyd, Tripo for inorganic
Character animation→ Runway Gen-4 for expressions, Kling for action i2v
Commercial product video→ Kling + SAM2 segmentation (the Garage Force pipeline)
Plasticine / claymation→ Minimax i2v with a Midjourney-styled input frame
Style / consistency→ IPAdapter with a trained Daechoo or Cyd reference
Inpaint / remove object→ SAM2 for mask, DALL-E 2 inpaint or Flux Fill for replacement
Pose-locked character→ Two-pass pose control: coarse sampler + ControlNet OpenPose refinement
Upscale to delivery→ Hi-res fix + ESRGAN for 4K, latent upscale for feathered detail
In Production

Two shipping creative projects built on this setup.

Cyd & Julie · Photo → Character → 3D Model → Animation

Animated IP · A real tortoiseshell cat becomes a rigged 3D character, animated across 25 expression clips
Cyd lucky cat render
Cyd render from the pipeline. Flux stylization with the Cyd LoRA.
Cyd pose test
Pose test — character consistency locked via IPAdapter.

Garage Force · SAM2 + Kling Pipeline at Scale

Commercial production · 115 real properties · 100+ transition effects tested. SAM2 isolates the floor, Kling i2v renders the transformation
Aurora Wave
Aurora Wave transition.
Fairy Dust
Fairy Dust transition.
Magic Snap
Magic Snap transition.
Power Wash
Power Wash transition.
Robot Squad
Robot Squad transition.
Wizard Spell
Wizard Spell transition.
Spring Bloom
Spring Bloom transition.
Cat Stampede
Cat Stampede transition.
Black Hole
Black Hole transition.
Golden Sweep
Golden Sweep transition.

Let's build something together

Available for editing, supervising, and tool-building consulting.

Contact
Ron Rauch