AI providers¶

Scelo's AI features (the scoped chats, model suggestions, narratives) run against an LLM provider. The default is local and private; hosted providers are opt-in.

Ollama (default — local, private)¶

By default Scelo talks to Ollama running on your machine: no key, no spend, no network. Your data never leaves the device. The default model is qwen2.5:7b-instruct.

Hosted providers¶

Settings → AI providers — the local default plus hosted, bring-your-own-key providers

Switch to a hosted provider in Settings → AI providers ("Bring your own AI"):

Anthropic (Claude)
OpenAI
Google Gemini
OpenRouter (one key, dozens of models)
Any OpenAI-compatible endpoint (set a base URL)

For each, set your API key, optionally a model, and (for compatible endpoints) a base URL. Re-saving with a blank field keeps your previous value, so you can update just the key without resetting the model.

Test connection¶

Each provider has a test connection button. It sends a tiny prompt and shows the reply, so you can confirm the key and model work before relying on them.

Reasoning models can return empty

Some models (e.g. gpt-oss, DeepSeek R1) spend their token budget "thinking" before producing visible text. Scelo gives the test a generous budget and falls back to the reasoning channel, but if you see "(connected — model returned no text)", the connection is fine — the model just needs more tokens or you can pick a non-reasoning model.

Reset¶

reset to defaults clears every saved key/model and switches back to the local Ollama default. Otherwise your settings persist across launches (keys are held in the OS keychain where available).

How it works in the desktop app¶

The desktop IDE ships no cloud backend. Chat and the "test connection" call the provider's HTTP API directly from the app's main process (no CORS, the decrypted key never touches the renderer). Replies fill in at once rather than streaming.

See also Chat for where these assistants appear.