Ollama

Ollama runs open-weight models locally with a small command-line server. The app talks to Ollama’s native API (/api/chat with a JSON-schema format, /api/embed), so structured output works without extra settings.

Verified with Ollama 0.40.7 on macOS (Apple Silicon).

1. Install a current version

Install from ollama.com/download or with Homebrew (brew install ollama), then check the version:

ollama --version            # 0.40 or later

On recent macOS releases, old Ollama builds do not work: 0.18 failed to initialize Metal and could not run models on the GPU; 0.40.7 works. If ollama run reports a Metal or GPU initialization error, upgrade before anything else.

2. Start the server with a models folder and enough context

OLLAMA_MODELS=~/Models/ollama OLLAMA_CONTEXT_LENGTH=32768 ollama serve

• OLLAMA_MODELS keeps Ollama’s blobs in their own folder. Do not share it with LM Studio’s models folder (LM Studio stops indexing folders with more than about 7,000 files).
• OLLAMA_CONTEXT_LENGTH sets the default context window. Ollama otherwise chooses 4K, 32K or 256K depending on available VRAM, and the app’s MSEL stage needs about 32K. You can instead set it per provider with the numCtx option (see Advanced provider options).
• The macOS menu-bar app starts its own server; quit it first if you want to run ollama serve with your own environment.

3. Pull models

ollama pull qwen3:30b-a3b-instruct-2507-q4_K_M   # chat model for generation (19 GB)
ollama pull nomic-embed-text                      # embedding model (274 MB)
ollama list

Any chat model Ollama serves can be used. The suggested tag is the Ollama build of the model the LM Studio reference runs use (Qwen3-30B-A3B Instruct 2507); it was not run through a full scenario on Ollama for this guide. Mixture-of-experts models like it give the best speed for their quality on a laptop; very small models (1B to 3B) answer the connection test and short stages but are not good enough to write a full MSEL.

4. Connect it in the app

1. Providers → Detect local services. Ollama appears with its model count. Click Add.
2. The detected provider’s default model is the largest chat model. The app asks Ollama what each model can do (capabilities in /api/tags, or /api/show on older servers), so embedding models such as nomic-embed-text or all-minilm are never picked or suggested for writing; if Ollama does not answer, names like embed, minilm, bge, e5, gte and mxbai are treated as embedding models. To use a different model, click Edit and set Default model (the exact name from ollama list, including the tag, for example qwen3:30b-a3b-instruct-2507-q4_K_M).
3. Click Test. Measured with qwen2.5:3b: “OK” in 21 ms with the model loaded, 11.7 s when Ollama first had to load it.
4. In Routing, assign Ideation, Drafting and Critique (or only some of them) to Ollama.

Manual setup: Add provider → Type: Ollama, Base URL http://127.0.0.1:11434.

To compare models on this computer, see Benchmarking local models; it sets num_ctx and a fixed seed for each request.

Ollama also exposes an OpenAI-compatible API at http://127.0.0.1:11434/v1. Prefer the native Ollama type: it passes the JSON schema and num_ctx directly.

Troubleshooting

Symptom	Fix
Detect says Ollama “Not found”	Start ollama serve (or the menu-bar app). curl http://127.0.0.1:11434/api/version should print the version.
Metal / GPU initialization error, or generation on CPU only	Upgrade Ollama (0.18 failed on recent macOS; 0.40.7 works).
model "…" not found	The app’s model name must match ollama list exactly, including the tag. ollama pull it first.
The first request is slow, later ones fast	Ollama loads the model on first use and unloads it after 5 minutes idle. Set OLLAMA_KEEP_ALIVE=1h when starting the server.
Long stages return truncated or invalid JSON	Context too small: start the server with OLLAMA_CONTEXT_LENGTH=32768, or set numCtx on the provider.
Models are stored in the wrong place	OLLAMA_MODELS must be set in the environment of the server process, not the shell you run ollama pull in.