Skip to content
Open Net Worth Docs

Models and cost control

Lucius isn’t tied to one AI provider. Bring your own API key (or point at a local model) and the app calls it directly — no markup, no proxy.

Choosing a provider and model

Settings → AI sets the instance default. Built-in presets: Anthropic, OpenAI, Google Gemini, xAI (Grok), Zhipu (GLM), OpenRouter, Ollama running locally, a local LiteLLM proxy, or a fully custom OpenAI-compatible endpoint for anything else those don’t cover. Already have Claude Code or Codex installed on this machine? Point at that CLI directly instead of an API key — it spends your existing subscription rather than billing per token. Your key is never sent to Lucius; the field hides it by default, but you (the operator) can reveal it to confirm what’s saved — logged.

A conversation or a saved Automation can pin its own provider/model independent of the instance default, so changing the default later doesn’t retroactively change an existing Automation.

Routing background tasks to a cheaper model

Several things happen in the background rather than as a direct reply — memory extraction/consolidation, news screening, a price-alert writeup, reading a scanned statement, and a few others. Each can point at its own model, independent of your main chat model: run conversation on a strong model, route high-volume, low-stakes work to something cheaper. Unset tasks use the instance default.

Context management

A long conversation is automatically compacted in the background once it gets large — older turns fold into a summary while recent exchanges stay intact, so it keeps working instead of hitting a hard limit or costing proportionally more forever. Nothing to configure. Separate from Memory: folding manages one conversation’s own size; memory persists across conversations.

Seeing what a turn cost

Every reply carries a small cost figure — what that exchange cost, in real provider pricing, and which model answered — so a surprisingly expensive reply is visible immediately, not buried in a monthly bill.

For the bigger picture, Settings → AI breaks spend down by time window (Today, 30 days, or All time), by surface, by provider, and by model, plus a projected monthly total extrapolated from the selected window’s rate.

  • lucius/memory — its own background-task model
  • lucius/automations — scheduled runs show their own cost