Chat

Every other node in NodeGPT is about building the pieces. This one runs the show. The Chat node is the pack's conversation driver - you point it at a user, point it at an agent, give it a task, and it fires off an AutoGen exchange between them. Everything upstream (the LLM providers, the agent personas) exists to feed this node, and everything downstream (DisplayString, Output2String, Conditioning) exists to catch what it produces.

The mechanics are the pack's heart: NodeGPT wrapped Microsoft's AutoGen into ComfyUI so you could choreograph multi-agent conversations as a graph, no Python required. The Chat node is a thin but revealing window into that design, because it asks you to manage the conversation by hand.

The inputs, and the weird one

• Task - the thing you want done. Think of it as the brief: "Write a Python script that draws a spiral."
• User - a user object. This is AutoGen's UserProxyAgent; in the sample workflows it represents the human side of the chat.
• Agent - your assistant (from the Assistant node, or any agent that produces an Agent output). This is who actually does the thinking.
• Instructions - a big multiline block with defaults that tell you how the node wants to be driven: NEW for a new chat, SEND to send a message, and a note to ask if there's a follow-up. This is the pack author teaching you the protocol.
• Input - the control channel. Type NEW to start a fresh chat, SEND to push the next message. This is the rawest part of the whole pack: you are manually steering the conversation loop by editing a text widget between runs.

Output: TEXT - the transcript or final reply, which you then wire into DisplayString to see it, Output2String to turn it into a plain string, or Conditioning to turn it into actual Stable Diffusion conditioning.

Why it behaves like this

AutoGen conversations are normally a loop that terminates on its own logic. In NodeGPT, that loop got exposed as a widget - which is exactly the quirk users hit in the wild. Reports from people running it describe the queue running away and "going to infinity," never terminating until ComfyUI is closed. The manual NEW/SEND protocol is the author's attempt to keep control, but the pack was openly unfinished, and the repo is now gone: xXAdonesXx/NodeGPT 404s today and the account is deleted, so the README's clone path is dead. If you load an old workflow using this node, expect to babysit the Input field and be ready to clear the queue.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/xXAdonesXx/NodeGPT

Restart ComfyUI so its requirements (including pyautogen) install on startup. Then wire LLM provider → Assistant → Chat, set Input to NEW, run once, then SEND to keep it talking. Just remember the historical asterisks: it's a fascinating artifact of AutoGen-in-a-graph, and a rough one to actually run in 2026.