Simple LLM Config: Ollama

The OpenAI, Claude, and Gemini config nodes all assume you've got an API key and are fine with your prompts leaving your machine. Simple LLM Config: Ollama is the exception in this pack, and for a lot of users it's the most interesting one: no key, no account, no data leaving your box. It points the workflow at a local Ollama server running on your machine, which is the private option for LLM stuff inside ComfyUI - the equivalent of running a local checkpoint instead of paying per image.

It's also the only config node that needs something installed beyond the pack itself: Ollama (from ollama.com) must be running and have a model pulled. The README's own framing calls out "Local LLM supports Ollama" - it's the pack's only local path.

How it works

The node builds an LLM_CONFIG with driver: "ollama", your model, a base_url, temperature, and max_tokens. When a prompt runs, the Ollama driver first hits {base_url}/api/tags to see what models are actually loaded, then POSTs to /api/generate with stream: false - no SDK, straight HTTP. Two nice details: it matches model names even when you type llama3 and the local copy is tagged llama3:latest, and it maps your max_tokens to Ollama's num_predict option.

The inputs that matter

• model - a free-text string, default llama3. This must match a model you've pulled, or the run errors out (see below).
• base_url - default http://localhost:11434. Leave it unless your Ollama runs elsewhere (another machine, a Docker container).
• temperature - 0 to 2.0, default 0.7.
• max_tokens - 1 to 16384, default 2000. Be aware local models are slow; this is the cap on generated tokens, not speed.

Output: a single LLM_CONFIG for Simple LLM Agent or Simple LLM: Agent with Rules.

Installing the pack - plus Ollama itself

# install the pack
cd ComfyUI/custom_nodes
git clone https://github.com/NakamuraShippo/ComfyUI-NS-Util
pip install pyyaml watchdog

Then install Ollama and grab a model - this is the step beginners miss:

# https://ollama.com/download, then:
ollama pull llama3

Restart ComfyUI after cloning the pack. ComfyUI 0.3+ recommended.

Common issues

The error that tells you everything: Error: Model 'x' not found. Available models: ... - the driver checks /api/tags before generating and lists what you actually have. Fix: ollama pull <name>. If you get connection errors instead, Ollama isn't running or the base_url is wrong - curl http://localhost:11434/api/tags is the fastest check. And set expectations on speed: a local model is private and free, but a 7B on CPU will make you miss the cloud. This whole LLM family is still marked in testing by the author, and Ollama is the newest surface, so treat early hiccups as part of the deal.