📎CrewAI LLM OpenAI

The name is a lie, and the lie is the useful part. 📎CrewAI LLM OpenAI doesn't actually talk to OpenAI by default - its defaults point at Ollama's OpenAI-compatible endpoint (http://localhost:11434/v1). What this node really is: a generic client for any OpenAI-compatible API. Set the base_url and it'll happily talk to Ollama, LM Studio, vLLM, Together, or OpenAI itself. It's the Swiss-army LLM node of this pack, and honestly the one I'd grab when I'm not sure which backend I'll end up on.

The inputs

• model - required, default crewai-llama3. That default is just whatever model the author was running when they wrote the examples; you'll replace it with your actual model name. For OpenAI proper, gpt-4o works. For Ollama, the model name is what you pulled (e.g. llama3.2).
• base_url - required, default http://localhost:11434/v1. This is the OpenAI-compatible endpoint. Ollama and LM Studio both expose one under /v1, so pointing here gets you local models with zero OpenAI involvement.
• api_key - optional, default pulled from the OPENAI_API_KEY environment variable. Most local servers don't care what you send, so you can often leave this blank or paste a dummy.

One output, llm (CREWAI_LLM), which feeds the 📎CrewAI Agent node like any other LLM node. The mechanism: LLM(model=model, base_url=base_url, api_key=api_key).

How to actually use it

• Point it at Ollama's /v1: keep the default base_url, set model to a pulled model like llama3.2. Free, local, same result as the dedicated Ollama node - the difference is mostly the name and the API surface.
• Point it at LM Studio: run LM Studio's local server, set base_url to http://localhost:1234/v1, model to whatever you loaded. People run CrewAI nodes in ComfyUI against LM Studio exactly this way and it works fine.
• Point it at OpenAI: set base_url to https://api.openai.com/v1 (or empty - CrewAI defaults to OpenAI), model to gpt-4o, and put a real key in api_key. Though if that's your plan, the 📎CrewAI LLM Chat GPT node does it more directly.

The one real gotcha is cost-adjacent, not setup: whichever endpoint you pick, the crew makes many sequential calls. A verbose multi-agent run against a paid API is where a "free" prototype turns into a surprise bill. Start cheap, add quality later.

Installing

Ships with ComfyUI-CrewAI. 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. Only dependency is CrewAI itself. And a tip: because this node is just an OpenAI-compatible client, it's the best fallback in the pack - if a backend-specific node (looking at you, Hugging Face) is broken, this one usually gets you going against the same endpoint.