Griptape Prompt Driver: Ollama

Ollama is the local-model driver, and it's the one the pack's own README walks you through first. No API key, no account, no per-token invoice - you install Ollama, pull a model, and a Griptape agent gets a free, private brain that lives on your machine. For learning how the whole agent thing works, or for any workflow where your data shouldn't leave your desk, this is the driver to start with. It's the reason the pack can do something useful on day one without touching a paid API.

The honest trade-off: "free" means "your GPU." Ollama runs models locally, competing with ComfyUI's diffusion for VRAM. A 7–8B model like Llama 3 is the sweet spot - small enough to coexist with image gen, capable enough for real agent work. Don't try to run a 70B and generate images on the same card.

How it works

Ollama runs a local server on http://127.0.0.1:11434. The node builds a Griptape OllamaPromptDriver pointed at it - and those two fields (base_url http://127.0.0.1, port 11434) are the only required inputs in this whole driver family, which tells you how foundational local is to the pack. The model you specify has to already exist locally (ollama run llama3 pulls it). Output is a single DRIVER.

The inputs that matter

• model - free-text, e.g. llama3, mistral, llama3.2-vision. It must be a model you've already pulled with Ollama; the node won't download it for you.
• base_url / port - the required defaults (http://127.0.0.1, 11434) match a stock Ollama install. Change only if you've configured Ollama otherwise.
• keep_alive - default 240. Seconds Ollama keeps the model warm in memory between requests. Lower it to free VRAM sooner, raise it to avoid reload latency on chatty agents. This one's actually useful.
• use_native_tools - defaults on here, unlike LM Studio; Ollama serves modern models that mostly handle tool calling fine.
• top_k (default 40) and min_p (default 0.1) - the sampling knobs; temperature defaults low (0.1) for reliability.

seed, max_tokens (-1 = model default), and max_attempts_on_fail complete the set.

Getting started (the README's own path)

# install Ollama from ollama.com, then:
ollama run llama3

That one command pulls the model and opens a chat to verify it works. Close it, restart ComfyUI, and the driver will find Ollama on the default port. The pack's YouTube tutorials cover this exact flow.

Installing the pack

ComfyUI Manager → Griptape → ComfyUI-Griptape, restart, or

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

No keys to set for the Ollama driver - the only "config" is having Ollama running.

Common issues

Most failures are "Ollama isn't running" - the server only exists while the Ollama app is up. Next most common: a model name in the field that wasn't pulled (model not found). And if ComfyUI breaks after installing the pack at all, it's griptape[all] stomping torch - reinstall with pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121. Stale griptape shows up as ImportError on driver names; python -m pip install griptape -U fixes it.