📎CrewAI LLM Chat GPT

This is the node that actually talks to OpenAI. It's the default in the pack's example workflows, and it's the one most people land on first - mostly because it's the only LLM node in ComfyUI-CrewAI that "just works" the moment you already have an API key. The others either need a local server running or, in one case, are broken as shipped.

The name is accurate for once: it wraps CrewAI's LLM object with your OpenAI credentials and hands that object down the wire. It does not generate images, it doesn't download a checkpoint, and it never touches your GPU. It's pure text-in, text-out plumbing that sits at the top of a crew. Nothing in this pack runs without one of the four LLM nodes somewhere in the graph.

What it gives you

One output, llm, of type CREWAI_LLM. That's the whole job: you feed it into a 📎CrewAI Agent node's llm input, and every agent in the crew thinks using that model. The chain you're building looks like this:

CrewAI LLM Chat GPT  →  CrewAI Agent  →  CrewAI Task  →  CrewAI Crew

Two fields, both strings:

• api_key - required in the schema, and its default is pulled from the OPENAI_API_KEY environment variable at node creation. If you've already exported that variable where ComfyUI runs, you can genuinely leave this blank. If not, paste a sk-... key here.
• model - optional, defaults to gpt-4o. Set it to whatever OpenAI model you want the crew to reason with.

Under the hood it's a one-liner: LLM(model=model, api_key=api_key). No temperature, no max tokens, no system prompt - CrewAI's agent takes care of all of that from the Agent node.

Installing it

The node ships in the ComfyUI-CrewAI pack by luandev. Easiest path is ComfyUI Manager - open it, hit Install Custom Nodes, search "ComfyUI-CrewAI", and install. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/luandev/ComfyUI-CrewAI
cd ComfyUI-CrewAI
pip install -r requirements.txt

Then restart ComfyUI. The only dependency is crewai (>=0.95, <1.0), which pulls a fairly heavy tree with it - langchain and friends - so give pip a minute. There are no model files to download: the actual "model" lives in OpenAI's cloud, and you pay per token.

Where people get burned

• Your key ends up in the workflow JSON. Whatever you type into api_key is saved with the workflow and embedded in saved images. Don't share a workflow with a live key in it.
• The default gpt-4o burns tokens on verbose crews. A multi-agent run with verbose on makes a lot of calls. If you're prototyping a prompt-refining loop, a smaller/cheaper model is fine and the output quality barely moves.
• It's synchronous. The whole crew executes inside ComfyUI's Python process, so the queue stalls while the agents chat. That's the pack's design, not a bug - just don't expect to queue up a dozen prompt-crafting jobs and walk away.

If you don't have an OpenAI key and don't want one, reach for the 📎CrewAI LLM Ollama node instead - same output, free, runs on your machine. This one is the right pick when you want the strongest reasoning with zero setup fuss.