Set Ollama LLM Service Connector 🐑

The MieNodes pack ships a family of Set*LLMServiceConnector nodes, one per LLM provider, and this is the local one. Want a small language model rewriting prompts, captioning images, or answering questions inside the graph - no subscription, nothing sent to a cloud API? This node is how you point all of that at your own Ollama install.

Get one thing straight before wiring anything: the name does extra work. Set Ollama LLM Service Connector doesn't call Ollama, doesn't generate text, and needs no API key. It's a configuration node in the ComfyUI sense - it packages host, model, and (irrelevant) token into an LLMServiceConnector handle for the pack's actual worker nodes. Its only output is llm_service_connector, which you feed into any node with an llm_service_connector input - Call LLM Service, the PromptGenerator family, the Kontext preset generator. The connector is plumbing, not the sink.

How it works

Mechanically it's simple and well-chosen. The node builds an object that posts to {host}/v1/chat/completions - Ollama's OpenAI-compatible endpoint, around since Ollama 0.3. Because Ollama ignores the Authorization: Bearer header (it only requires it to be non-empty), the node sends a placeholder token and api_token can stay empty or hold any junk. Leave model blank and it silently falls back to qwen2.5.

That's the "external server" pattern: Ollama runs as its own process and ComfyUI just talks HTTP to it - the opposite tradeoff from in-graph GGUF loaders that keep the model inside ComfyUI's VRAM budget. Two processes, and the diffusion model and the LLM share your card. Fine for prompt enhancement on a local 8B; don't expect a 32B and a Flux checkpoint to coexist happily.

The inputs that actually matter

Two fields to set, the rest ignored:

• model - the name of a model you've actually pulled on the Ollama host (qwen2.5, llama3.2, llava, deepseek-r1, qwen2.5vl:7b). Free text, no dropdown. Empty = qwen2.5.
• host - the bare Ollama URL, default http://127.0.0.1:11434. Ollama on another LAN box? http://192.168.x.x:11434, with OLLAMA_HOST=0.0.0.0 set on the server. Ollama in Docker, ComfyUI in another container? http://host.docker.internal:11434. No trailing slash, no /v1/... - that's appended for you.

The rest are mostly "set once, forget." api_token is a placeholder as covered above, and config_file, config_key, prefer_local_config are the pack's shared JSON-key plumbing - for Ollama there's no secret to store, so ignore them. The one worth knowing is timeout, default 60s. Ollama cold-starts a model from disk into VRAM on first call - 30–90 seconds for a 7B+ - so if you run big models and the first request keeps timing out, bump it to 120–180s. The base class retries on timeout; a longer single-attempt timeout just stops you wasting the first try.

Installing and wiring it

Install the pack the normal way: ComfyUI Manager → search "MieNodes" (the pack title is ComfyUI_MieNodes) → install, then restart. Manual route is the same as always:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes

Then restart ComfyUI. The pack's requirements.txt pulls a grab-bag of file/image utilities (opencv, imagehash, huggingface_hub, soundfile and friends) that its other nodes need - let it install. Ollama itself is a separate install; this node talks to it over HTTP, so make sure it's running with a model pulled:

# on the Ollama host, once
ollama pull qwen2.5

In the node graph: add Set Ollama LLM Service Connector (under 🐑 MieNodes/🐑 LLM Service Config), set the model name, add Call LLM Service, connect llm_service_connector → llm_service_connector, type a prompt into input_text, and queue. The pack's workflows folder has ready-made examples of the connector feeding the prompt-enhancer and captioning nodes.

Gotchas worth knowing

• Connection refused on the call node almost always means Ollama isn't running or host is wrong. Open http://127.0.0.1:11434 in a browser - no answer there, no answer here.
• Model not found is a model typo or a model you never pulled. Fix with ollama pull <name> on the host.
• First call hangs, then dies - that's cold start. Raise timeout. Ollama-level knobs like num_ctx and keep_alive aren't exposed through the OpenAI-compat endpoint; tune those in a Modelfile on the host.
• Vision models work (llava, qwen2.5vl): connect an IMAGE to Call LLM Service's image input and it's forwarded as inline data.
• Reasoning models (deepseek-r1, qwen3) emit <think>…</think> chains before answering; the pack strips them, so no regexing them out of your prompt.

MieNodes is a small, content-creator-maintained pack that touches your network by design, so the usual custom-node rule applies: it's open source and on the registry, but glance at what you're running before wiring a fresh pack into a serious workflow. For the actual use case, this is about the cheapest prompt-enhancement setup you'll find - free, offline, uncensored, and it never asks for a key.