📎CrewAI LLM Ollama

This is the node I'd reach for first. The pack's whole pitch is "let a crew of LLM agents write your image prompts," and this is the only LLM node in it that runs entirely on your machine - no API key, no cloud account, no per-token bill. If you already run Ollama for anything else, this node is a two-minute setup. If you don't, the payoff is the same for a one-time ollama pull.

What it does

It builds a CrewAI LLM from two inputs and passes it to your agent:

• model - required, default ollama/llama3.2:3b. Note the ollama/ prefix; that's CrewAI's way of saying "this is an Ollama model." Any model you've pulled into Ollama works here (ollama/llama3.1:8b, ollama/qwen2.5:7b, ...).
• base_url - required, default http://localhost:11434, which is where Ollama listens. Change it only if your Ollama runs on another machine or port.

Output is one llm (CREWAI_LLM), wired straight into the 📎CrewAI Agent node's llm input. The mechanism is a one-liner - LLM(model=model, base_url=base_url) - and it's the same downstream as every other LLM node here. It does not call an OpenAI-style API and needs no key, which is exactly the point.

Getting it running

Two steps before this node does anything useful:

ollama pull llama3.2        # or whatever model you want to use
ollama serve                # keep it running in the background

Then drag in 📎CrewAI LLM Ollama → 📎CrewAI Agent → 📎CrewAI Task → 📎CrewAI Crew, set a topic, and run. The pack's simple_crew.json example is basically this, minus the image sampler.

The honest trade-offs

The default llama3.2:3b is a 3-billion-parameter model. It works, and it's a fine way to learn the pack - but for genuinely good prompt writing you'll want the 8B or larger. The 3B tends to produce flat, generic prompt pairs. The catch is that bigger local models need VRAM, and a multi-agent crew makes many sequential calls, so the run gets slow fast.

Two things that trip people up:

• The whole run is synchronous. The crew executes inside ComfyUI's Python process and the queue freezes while agents chat. On a local model that can be a while. This is the pack's design, shared by all its LLM nodes - not an Ollama bug.
• Ollama needs to be reachable. If base_url is wrong or Ollama isn't running, you get a connection error at the Crew node, not at this one - the LLM isn't actually called until the crew kicks off.

Installing

The node ships in ComfyUI-CrewAI (by luandev). Install via ComfyUI Manager (search "ComfyUI-CrewAI") or:

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

Restart ComfyUI. The only pip dependency is CrewAI itself - no diffusion models, no tokens. If you want the strongest free reasoning without a local model, the 📎CrewAI LLM OpenAI node pointed at any OpenAI-compatible endpoint is the sibling to reach for.