Ollama

A local LLM server. Lens writes and refines captions through it. Darkroom can expand prompts with it. The SingGuard guard server runs its model on it.

Used by

Package	What for	Reads
argus-lens (opens in new tab)	The `openai-compat` caption backend and the refine pass. Both default to Ollama's OpenAI-compatible endpoint on localhost.	`ARGUS_OPENAI_COMPAT_BASE_URL`, `ARGUS_OPENAI_COMPAT_MODEL`, `ARGUS_REFINE_BASE_URL`, `ARGUS_REFINE_MODEL`
argus-darkroom (opens in new tab)	Optional `--expand` prompt expansion. Uses Ollama's native `/api/chat`, so vLLM and LM Studio won't work here.	`ARGUS_DARKROOM_OLLAMA_URL`, `ARGUS_DARKROOM_OLLAMA_MODEL`, `--expand`
SingGuard guard server	Loads the SingGuard GGUF from Hugging Face. See SingGuard.	`OLLAMA_URL`, `GUARD_MODEL`
argus-studio (opens in new tab)	The `lens` and `full` profiles run a bundled Ollama and pull lens's refine model.	`OLLAMA_TAG`

Official guides

Minimal setup

Install Ollama (native or Docker) and pull the models you'll use. For example, lens's refine default and a vision model for captions:
```
ollama pull qwen3:8b
ollama pull qwen3-vl:8b
```
Ollama listens on 127.0.0.1:11434 by default. For containers or other machines to reach it, set OLLAMA_HOST=0.0.0.0 on the Ollama side, and only on a trusted network: Ollama has no authentication.
Point the clients at it:
- lens uses the OpenAI-compatible path, with /v1: ARGUS_OPENAI_COMPAT_BASE_URL=http://<GPU_HOST>:11434/v1 and ARGUS_OPENAI_COMPAT_MODEL=qwen3-vl:8b. From inside a container on the same machine, use http://host.docker.internal:11434/v1.
- darkroom uses Ollama's native API, with no /v1: ARGUS_DARKROOM_OLLAMA_URL=http://<GPU_HOST>:11434 and a model in ARGUS_DARKROOM_OLLAMA_MODEL. Both must be set for --expand to work.

Studio's lens profile runs its own Ollama, and a one-shot ollama-pull fetches the refine model, so you need neither step for that profile.