VibeComfy x Hivemind: Agents That Read and Edit ComfyUI Workflows

ComfyUI Wikinews

VibeComfy translates ComfyUI workflow JSON into editable Python for CLI agents, and now pairs with the Hivemind knowledge base for community-sourced workflow patterns.

VibeComfy mascot

VibeComfy turns ComfyUI workflow JSON into Python that coding agents can read, edit, validate and compile back.

A translation layer between agents and workflows

VibeComfy by peteromallet reached a milestone the author considers robust enough for daily use: workflows become editable Python, and agents gain access to Hivemind, an open knowledge base for the ComfyUI ecosystem. The update works both inside the ComfyUI editor and through the command line on top of pip-and-uv-installable ComfyUI, and the author reports it runs with existing Claude Code, Codex or Hermes local credentials.

The core idea is a translation layer called VibeWorkflows. VibeComfy imports a ComfyUI workflow, represents it as ordinary Python (a build() function calling typed node wrappers), validates it, and compiles it back to the API JSON that ComfyUI queues. JSON remains the import and export format; Python is the authoring surface. The reasoning: JSON graphs are hard for small local agents to reason about or validate, while Python is the language even lightweight agents handle best.

How it differs from existing bridges

ApproachFocus
ComfyScriptIDE-oriented Python export, no longer maintained against the modern ecosystem
ComfyMCPProtocol layer for operating ComfyUI
VibeComfyAuthoring layer: understanding workflows, finding proven patterns, making complex edits, preserving results

Generated template files are marked read-only (# vibecomfy: generated) and meant to be copied into a gitignored recipes/ workspace before editing, so agents can mutate a personal copy without destroying the proven pattern.

Getting started

VibeComfy installs into an existing ComfyUI checkout as a custom node (with an in-editor agent panel), or runs standalone for CLI agents. Setup instructions are written to be pasted directly into your agent of choice, and the author is explicitly asking for feedback on broken installs and confusing steps, plus help hardening the project's agent eval suite.

Comments

Sign in with GitHub to join the discussion.

Loading comments…