Your own ComfyUI workflows and LoRA training on RunComfy's cloud GPUs.
The whole RunComfy loop is covered, from deploying a ComfyUI workflow as an API and running it or a hosted model to training a LoRA on your images and trying the result.
What it is
RunComfy's own remote MCP server at mcp.runcomfy.com/mcp. It works with your RunComfy account, so you see only your own deployments and all usage is billed to you. It covers three RunComfy products: ComfyUI deployments, hosted models and LoRA training.
What you get
- ComfyUI workflows deployed as serverless APIs, with hardware and autoscaling chosen, then changed or deleted
- Jobs sent to a deployment, checked on, and their output files fetched, or cancelled to stop GPU use
- Hosted models browsed by keyword, capability or price, and run without any deployment
- Training datasets created from files or captions, and LoRA training jobs started, resumed or cancelled
- A trained LoRA passed straight to a hosted model to try it
- Your remaining RunComfy balance
Requirements
A RunComfy account with a balance to spend. You connect with the API token from your RunComfy Profile, or by signing in through your browser in clients that support it. RunComfy gives setup steps for Claude Code, Claude.ai, ChatGPT, Cursor, VS Code and Windsurf. Runs and training use your RunComfy balance, and a deployment kept warm with a minimum instance count is billed even when idle, so check the balance and cancel jobs you no longer need.
Setup effort
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen