Connecting a provider

A provider supplies the models the agent runs on. Smart Second Brain is provider-agnostic: you bring your own, and you can configure several at once and switch between them per chat.

Add one under Settings → Smart Second Brain → General → Providers → Add provider.

Supported providers

Provider	Runs locally	Notes
Ollama	Yes	Fully local. Nothing leaves your machine.
oMLX	Yes	Local inference on Apple Silicon.
OpenAI	No	API key, or ChatGPT sign-in.
Anthropic	No	Claude models.
OpenRouter	No	Routes to many vendors behind one key.
Custom (OpenAI-compatible)	Depends	Any endpoint speaking the OpenAI API.

You can create multiple instances of the same provider type with distinct display names and endpoints, useful for separating a work key from a personal one, or pointing at two different self-hosted endpoints.

Local setup with Ollama

1. Install Ollama and start it.
2. Pull a chat model and an embedding model:

ollama pull llama3.1 && ollama pull qwen3-embedding:0.6b

3. In plugin settings, add an Ollama provider. The default base URL is http://localhost:11434.

Local providers are treated as trusted by default, since no data leaves the machine.

Other local engines

LM Studio, a llama.cpp server, or anything else that speaks the OpenAI API works through the Custom provider (the OpenAI-compatible template). For LM Studio, start its local server and set the base URL to http://localhost:1234. The API key can be left empty.

A Custom provider starts untrusted, because the plugin cannot tell a local endpoint from a remote one. If you want it to see private notes, mark it trusted in the provider’s settings. See Trusted providers.

Cloud setup

Add the provider, paste your API key, and select a model. Keys are held in Obsidian’s secret storage. The plugin’s config file stores only a reference, never the raw key.

Cloud providers are untrusted by default, which means they are blocked from reading, embedding, or receiving any note you’ve marked private. See the Privacy model.

Choosing a model

Use the most capable model your provider offers. Agent quality depends heavily on reasoning and tool-use ability. Frontier cloud models give the best results. For a fully local setup, pair a capable Ollama chat model with a strong embedding model.

Model names move faster than this page does. For a current comparison, Artificial Analysis ranks chat models on quality against price and covers open-weights models alongside hosted ones. Weight its agentic and tool-use benchmarks over the headline intelligence score — the agent spends most of its time calling tools, and a strong reasoner that calls them unreliably is the worse choice here.

For embedding, qwen3-embedding is a good default. It is multilingual, and it is available on Ollama, oMLX and OpenRouter, so the same choice follows you between a local and a hosted setup. Sizes go 0.6b / 4b / 8b. 0.6b is the laptop-friendly one.

To compare alternatives, the MTEB leaderboard is the standard reference. Read the Multilingual board rather than the English one unless your vault is English-only, and compare within a size group — it buckets models at <500M, 500M–1B, 1B–5B and >5B, and the overall top rows are large hosted models that a local setup can’t run. What matters for retrieval quality is the best model in the size you can actually afford to run.

Note

Some embedding models expect a search query to be phrased differently from the text being searched. The plugin handles that automatically for the families that need it — Qwen3-Embedding, harrier, gte-Qwen2, BGE English and mxbai — based on the model id. There is no setting, and switching models needs no reindex. See Query instructions.

Tip

Vision and PDF support is detected per model at runtime, so a model that supports images will accept them without extra configuration.