Ollama Connectivity

Think of this node as a checkpoint loader, except instead of loading weights into ComfyUI's own memory, it's pointing at a server that already has them. OllamaConnectivityV2's only job is connection: which Ollama server, which model on it, and how long to keep that model warm. Everything else in the pack's current node family - OllamaGenerateV2, OllamaChat - plugs into this one instead of repeating the same server address five times across your graph.

This matters because comfyui-ollama is a client, full stop. It doesn't ship a model, doesn't run inference locally, and doesn't bundle Ollama. You need a running Ollama server reachable from wherever ComfyUI is running - could be the same machine, could be a box down the hall - and this node is the thing that finds it. Its own description says as much: "provides connection to an Ollama server. Use the refresh button to load the model list in case of connection error or after installing a new model." That refresh button is not decoration; it's the actual fix for most problems with this node.

The inputs that matter

• url - defaults to http://127.0.0.1:11434, Ollama's default local port. Change this if the server's on a different machine, a different port, or inside a container.
• model - a dropdown populated live from whatever's actually installed on that server. If it's empty, either the server isn't reachable at that URL, or there are no models pulled yet.
• keep_alive (-1 to 120) and keep_alive_unit (minutes/hours) - how long Ollama keeps the model resident in memory after a call finishes. -1 means indefinitely, 0 unloads it immediately. This isn't cosmetic: a model sitting loaded is VRAM you don't get back for your diffusion model. If you're tight on GPU memory, either keep this short or move Ollama to its own machine - people running heavy SD workflows alongside an LLM do this specifically to keep VRAM free.

Output: a single connection socket (type OLLAMA_CONNECTIVITY) that any Generate or Chat node's connectivity input accepts.

Installing it

Via ComfyUI Manager, search comfyui-ollama (by Stav Sapir), or ollama and pick the stavsap one - it's the most commonly referenced Ollama integration for ComfyUI, though a couple of alternatives exist if you go looking (IF_AI_tools, LLM Polymath) with slightly different node shapes. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/stavsap/comfyui-ollama
pip install -r requirements.txt

then restart ComfyUI. This node's dependency footprint is small - the heavy lifting (actually installing and running Ollama) happens outside ComfyUI entirely: curl -fsSL https://ollama.com/install.sh | sh on Linux, the desktop installer on Mac/Windows, or a Docker container (CPU or --gpus=all). Pull at least one model before expecting this node's dropdown to show anything.

Common issues

Empty model dropdown. Either the server isn't running, or it's not reachable at the url you set, or nothing's been pulled yet. Hit the node's own refresh - it's specifically there for this.

"Value not in list: model" on a loaded workflow. The dropdown gets baked into the saved graph. Open a workflow someone else made, or one you made before pulling different models, and the saved model name won't match your current list. Refresh, then reselect.

Ollama Cloud / authenticated models. If you're pointing at a cloud-hosted model template that requires auth, this node has no key field - you set that up on the Ollama side instead, either by running ollama signin or manually placing your public key at ollama.com/settings/keys. The key file itself lives at a fixed OS path (~/.ollama/id_ed25519.pub on macOS, /usr/share/ollama/.ollama/id_ed25519.pub on Linux).

Container/remote setups. 127.0.0.1 only works if Ollama is on the same machine as ComfyUI. Docker or a separate host means pointing url at the real reachable address.