Ollama Client

Ollama Client is the socket that holds the whole local-LLM half of comfydv together. It doesn't generate anything - it just says "here's where my model server lives" and hands the rest of your graph a client output you can wire into the model selector, the load/unload nodes, and Chat Completion. It's the boring part, and that's the point.

Why you'd reach for it

You can't do any of the chat stuff without one of these sitting upstream. Think of it as the "connection" node: drop it on the canvas, make sure the endpoint is right, and every LLM node below it stops caring about the backend. The neat part of comfydv is that this node and the LlamaCpp Client node output the same LLM_CLIENT socket type. So switching from Ollama to llama.cpp - or back - is literally rewiring one client node; nothing downstream knows or cares.

How it works

There's one input that matters: endpoint, defaulting to http://localhost:11434, which is exactly where a stock Ollama install listens. That's all you need for a local setup. If your server lives elsewhere (a second machine on your network, a Docker container), point it there. The node wraps the connection in a provider object - Ollama's native REST API under the hood - and emits it as the client output.

A note if you're following along with older docs: the current shipped code also exposes an optional headers socket (OLLAMA_HEADERS) on this node, which is where comfydv's Ollama Header - nodes chain in if you need auth against a remote server. The socket is empty by default, so local installs can ignore it entirely.

The workflow shape

Typical graph:

1. Ollama Client → 2. LLM Model Selector (pick a model) → 3. Chat Completion (prompt it) → optionally LLM Load Model before and LLM Unload Model after for explicit VRAM control.

Install

ComfyUI Manager (search "comfydv") or:

cd ComfyUI/custom_nodes
git clone https://github.com/darth-veitcher/comfydv.git

Restart ComfyUI, then bring your own backend: ollama serve on the machine, ollama pull qwen2.5:latest to grab a model. Runtime deps (jinja2, pydantic, pydantic-ai-slim) install automatically.

Gotchas

• The dropdown is empty? The model list is populated when ComfyUI starts. Start Ollama before ComfyUI, or click the ⟳ Refresh button on the selector/load node after pulling a new model. Models pulled after server start won't show until you refresh or restart.
• Wrong endpoint = a specific, actionable error, not a silent empty dropdown - that's a deliberate design choice in this pack, so read the message.
• Keep it local. Anything that lets ComfyUI reach out to a network server deserves a second look at where it's pointing. This one's default localhost is the safe answer.