Assistant

When someone says "just have an AI do it," this is the node they're picturing. The Assistant node in NodeGPT is the AutoGen AssistantAgent with training wheels: you hand it a brain, give it a name and a system message, and it becomes a conversational agent that can hold up its end of a multi-agent chat. It's the pack's default, the agent you start with, and the one you'll wire up in almost every workflow before you add specialists.

NodeGPT was a 2023 experiment that pulled Microsoft AutoGen into ComfyUI, and its mental model is: LLM providers produce an LLM object, agent nodes wrap that object in a persona, and the Chat node actually runs the conversation. The Assistant is the generic persona. Its siblings (Critic, Engineer, Planner, Executor) are the same machinery wearing different hats.

The inputs that matter

• LLM - required. Comes from any provider node: ChatGPT, LM_Studio, Ollama, llama-cpp, oobaboogaOpenAI. No LLM, no personality.
• name - defaults to assistant. This is what shows up in the chat transcript, so name it something meaningful if it's one of several agents talking.
• system_message - defaults to You are a helpful assistant. This is the entire personality; make it do real work. "You are a Python expert. Always answer with code only." beats the default every time.
• Seed and Temp - both default to 0, which in AutoGen-speak means deterministic-ish and greedy. Bump Temp if you want variety; leave it at zero when you want the same answer twice.
• request_timeout - 120 seconds by default; if you're on a slow local model, this is the thing to raise when calls silently fail.

Output: Agent, ready to plug into Chat, GroupChat, or AppendAgent.

Wiring it up

The minimal working graph is: LLM provider → Assistant → Chat, with a User input on Chat and some Input text like NEW to kick things off. For bigger ambitions, feed the Assistant into GroupChat alongside an Engineer or Critic and let them hash it out.

Where it stumbles

Two pack-wide quirks you should expect, both reported by people who actually ran this in the wild. One: agents sometimes just echo back whatever was passed to them instead of generating - the fix people landed on was adjusting the Temp/Seed combo and checking the model preset on the server side. Two: a chat can spin without terminating, and you'll be killing the queue (or the whole app) to stop it. The author flagged the pack as unfinished, and the repo is now deleted - xXAdonesXx/NodeGPT 404s today, so the install path below is historical.

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

Restart ComfyUI so requirements (including pyautogen) auto-install, and you're off. For today's workflows, the concept - LLM config, persona, conversation driver - is worth understanding even if you end up building it in a maintained pack instead.