ComfyGen — Development skill for Claude Code

comfy-gen

**An agentic CLI for running ComfyUI workflows on RunPod serverless.** Designed to be used by AI coding agents (Claude Code, Cursor, Codex, Windsurf, Gemini CLI) as much as by humans. All output is structured JSON. No interactive prompts. Every command has verbose `--help` so agents can discover capabilities without documentation.

Table of Contents

• Use with Your Favorite AI Agent
• How It Works
• Installation
• Quick Start
• Commands
  • submit
  • status
  • cancel
  • download
  • config
• Getting Models onto Your Volume
• Configuration
• Storage
• Workflow Format
• Output Format
• Prerequisites

Use with Your Favorite AI Agent

ComfyGen ships with an **agent skill file** at `SKILL.md` in the repo root. Drop it into your agent's skill/tool directory and it can submit workflows, download models, check job status, and manage your RunPod infrastructure — all through natural language.

You: "Submit this workflow with seed 42 and download the output"
Agent: runs comfy-gen submit workflow.json --override 7.seed=42, parses JSON, fetches URL

Works with any agent that can run shell commands and parse JSON. The skill file teaches your agent the full API surface, including workflow analysis, parameter overrides, and error handling.

How It Works

comfy-gen submit workflow.json

1. Detects local file references in your workflow and uploads them to S3
2. Submits the workflow to your RunPod serverless endpoint
3. Polls for completion with real-time progress
4. Returns a JSON result with output URLs

Workers spin up on demand, execute the workflow, and shut down. You pay