Providers

A provider is where a model call goes. Sunday runs on Ollama by default, so nothing leaves your machine. You can send some roles to a hosted model and keep the rest local.

Roles

Role	What it does	Providers
Chat	Replies, quizzes, grading, skills	any
Planner	Picks the next step when Sunday acts on your computer	any, follows Chat until you set it
Vision	Reads diagrams and non-text screens	Ollama only
Embeddings	Memory and note search	Ollama only
Router	Decides what an utterance means	built-in rules, no model

Set one up

Open Settings → Models and press Change on Chat or Planner. Choose a provider, paste its key if it needs one, pick a model, then press Test and use. Nothing is saved until the test passes, so a wrong key or model shows up there, not mid-sentence. Switching the planner to a hosted model while chat stays on Ollama takes one pass through that form.

Keys live in your operating system's keychain, never in the database, the repo or the logs. Sunday doesn't show a saved key again; you can replace or remove it.

Prefer environment variables? They win over saved settings, which is handy for CI and scripts.

Provider	Key variable	Other settings
ollama	none	runs on localhost:11434
openai (OpenAI, OpenRouter, vLLM, LM Studio, llama.cpp server)	OPENAI_API_KEY, optional for local servers	OPENAI_BASE_URL
anthropic	ANTHROPIC_API_KEY
gemini	GEMINI_API_KEY	uses Google's OpenAI-compatible endpoint
nova (Amazon Nova on Bedrock)	AWS_BEARER_TOKEN_BEDROCK	AWS_REGION, default us-east-1; replies arrive in one piece, not word by word

SUNDAY_PROVIDER and SUNDAY_MODEL still pick the provider and model for chat when nothing is saved in Settings.

Privacy

Anything but Ollama sends the prompt to that provider. A prompt can include your notes and memories that matched the question, and the text of web pages or the screen when you asked Sunday to look at them. Settings shows a warning next to every role that leaves your machine. Embeddings and vision stay on Ollama, so notes and screenshots are never sent for indexing.

Cost

Sunday sends short prompts with a few retrieved notes, and the planner makes up to six calls per task. Hosted prices change, so check your provider's page. The cheapest way to try a hosted model is the planner role alone.

Test a provider

SUNDAY_PROVIDER=openai OPENAI_BASE_URL=http://localhost:11434/v1 SUNDAY_PROVIDER_TEST=1 \
  cargo test conformance::live -- --nocapture

This runs a real request and a real stream against the chosen provider. The same checks run in CI for every provider against a local mock server.

Add a provider

1. Implement Provider in src-tauri/src/provider.rs: complete, stream, models, warm, name.
2. Add it to build and to the provider list in Settings.
3. Add a mock that answers in its wire format and a test that calls check, as the existing providers do.
4. Keep the safety rule: a new provider is opt-in and its calls only happen after the user picks it.