enidpinxit/pinxit-comfycloud-mcp-codex

This repository is a set of skills and setup instructions for connecting OpenAI Codex to Comfy Cloud, a hosted service that runs AI image, video, audio, and 3D generation workflows. Comfy Cloud uses a protocol called MCP (Model Context Protocol) to expose its tools to AI agents, and this package teaches Codex how to use those tools effectively. The package includes two installable skills. The first covers the full Comfy Cloud generation workflow: submitting jobs, uploading media, polling for results, cancelling jobs, and retrieving outputs. The second is a companion skill focused on understanding ComfyUI workflow graphs and handling the different JSON formats that ComfyUI uses for its visual node editor versus its API. Together they give Codex enough context to plan and execute multi-step AI generation tasks without hard-coding assumptions about which models or templates are currently available. Setup is straightforward on Windows: copy the skills into your Codex skills directory, add the MCP server URL and your Comfy API key to Codex's config file, and restart Codex. The repository includes an example config snippet and step-by-step documentation covering the setup and known caveats around authentication. The repo also includes results from two real test runs: one that generated a 120-second Berlin-style techno track using the ACE-Step audio model, and one that recreated an image using the Krea2 Turbo model. These serve as proof that the setup actually works end to end, not just in theory. This is an unofficial companion package, not an official Comfy Org project. It was built from real testing sessions. The authoring rule for the skills is intentional: they teach Codex how to reason and search, rather than hard-coding lists of models or templates that would go stale as Comfy Cloud's catalog changes. The license is not mentioned in the README.