AI Chat

The name makes it sound like a chat box you type prompts into, and the first thing to know is that it isn't. Drop an AI Chat node on the canvas and you get a small button labelled "AI Agent". Click it and a floating dialog opens where you can tell the pack what to do in plain English - "lower all LoRA weights to 0.5", "switch to the cyberpunk style LoRA at 0.8" - and it goes and does it to your workflow. It's a conversational remote control for the LoRA stack, not a second brain that writes prompts for you.

Here's the honest part: this is a stub node. The backend execute() does nothing and returns nothing. All the work happens in the frontend, which is why it has no inputs and no outputs - it's an output node in the Moton PromptCraft category, and its only job is to be the door.

How it actually works

When you drop the node in, the pack's frontend hooks the canvas and adds that "AI Agent" button. Clicking it does two things: it finds the first Model & LoRA Group Loader in your graph, then opens the Agent floating panel pointed at it. From there the panel sends your message to whatever OpenAI-compatible API you've configured and a hardcoded system prompt parses your request into one of eight structured operations - lora_add, lora_remove, lora_toggle, lora_weight, checkpoint, prompt_set, category_set, and query. It's the same "agent translates natural language into structured operations" pattern the LLM-in-the-graph crowd has been converging on, just scoped to this pack's own LoRA and base-model system.

Two things follow from that. First, it needs your own LLM API key - the README recommends DeepSeek / GLM / Qwen, configured under ComfyUI Settings → PromptCraft. No key, and the chat has nothing to answer with. Second, the agent's reach is deliberately narrow: it can only poke at the LoRA stack and the base model, and it can't rewire the graph. Want it to swap a sampler or connect nodes? Not yet.

When it's worth it

Honestly, this is the sort of node you reach for after you've built a real workflow around the pack's LoRA system. If you already keep a Model & LoRA Group Loader stack in every workflow and you're tired of opening the stack editor to nudge weights, "drop everything to 0.5" is genuinely faster than hand-editing five sliders. If you're just trying to stack LoRAs and don't care about the rest of PromptCraft, you're better off with a text box and the community's battle-tested loaders - don't install this pack for this node alone.

Install

ComfyUI Manager → search ComfyUI-PromptCraft, or:

cd ComfyUI/custom_nodes
git clone https://github.com/moton16/ComfyUI-PromptCraft

Then restart ComfyUI. No model files to download; the only setup is the optional LLM key in Settings.

Troubleshooting

• The chat does nothing or errors. Check your API URL, key, and model in Settings → PromptCraft, then hit the "Test API Connection" button. The pack's own FAQ points here first.
• The agent says it can't operate on anything. It's looking for a Model & LoRA Group Loader node in the graph. No loader, no operations - you get a chat with nowhere to go.
• Don't paste secrets into the panel thinking they're local. The API key lives in a config file under your ComfyUI user dir and is masked in the settings UI, but the calls go to whatever API URL you configured.

One more honest note: this pack has essentially zero community footprint - no Reddit threads, no noise. Read the source or the README before trusting it with your API keys, same as you would any fresh LLM-enabled node.