Models and cost control

Lucius isn’t tied to one AI company. You connect your own provider — a Claude or ChatGPT subscription you already pay for, an API key, or a model running on your own computer — and the app talks to it directly. Open Net Worth adds no markup; you pay your provider, or nothing with a local model.

Set up a model

Open Settings → AI.
In AI providers, either:
- under Connect a subscription, follow the steps for Claude Code CLI or Codex CLI (install it, then sign in from a terminal) — no API key needed; or
- click ＋ Add provider, choose the Provider, paste your API key (or a Base URL for a local or custom endpoint), pick a Default model, and click Add provider.
Under Default model, choose the model Lucius should use unless a chat or Automation picks another.
Check it works: ask Lucius “what’s my net worth?” in a new conversation.

Step-by-step instructions for each provider, and where to get a key, are in Connecting an AI provider.

The providers you can add: Anthropic, OpenAI, Google Gemini, xAI (Grok), Zhipu (GLM), OpenRouter, Ollama (local), LiteLLM Proxy (local), Custom (OpenAI-compatible) for any other endpoint that speaks the OpenAI API, and the Claude Code CLI, Codex CLI and Antigravity CLI programs if they’re installed on this computer.

Each provider row shows whether it works: Ready, Key rejected or Can’t connect (for a CLI: Installed or Not installed). Click Edit on a row to change it, Test connection, or Remove it. The switch on a row turns a provider off without deleting it.

Your API key is stored on this machine and never sent to Lucius or shown in a chat. The key field hides it; you can reveal a saved key to check it.

Choose a model for one chat or one Automation

A chat: click the model name in the message box’s toolbar (it shows Default until you change it) and pick another. That conversation keeps it.
An Automation: in its editor, set Model under the instructions. Changing your default later doesn’t change an Automation that has its own model.

Preferred models (in Settings → AI) is your shortlist: the models you pick there are pinned to the top of every model picker.

Effort

Effort is how hard the model thinks before answering — more effort is slower and costs more, less is quicker and cheaper.

In a chat: open the model picker and use its Effort tab. The choices depend on the model: Low, Medium, High, Xhigh, Max, plus Off on some OpenRouter models. A new chat starts on Medium. Models that don’t support effort don’t show it.
In an Automation: Low effort, Medium effort or High effort, beside its Model.

Send background jobs to a cheaper model

Some of Lucius’s work happens in the background rather than as a reply. In Settings → AI → Task models each job can use its own model; anything you leave on Default uses your default model. Chat itself always uses your default (or the chat’s own choice).

Group	Tasks
Assistant	Memory, Chat suggestions, Conversation titles, Image reading (needs a model that can see images)
Insights	Intel & watchlist, Daily briefing (the morning note on the Dashboard), News screening, Alert analysis, Price alerts
Statement import	Statement import — reading statements end to end

A good pattern: a strong model for chat and Alert analysis, a cheap fast one for Conversation titles, Chat suggestions and Memory. Reset puts every task back on your default. If a task that reads images is set to a text-only model, a warning tells you that those runs will fail until you choose a model with vision.

The Daily briefing Automation is a different thing — it runs on that Automation’s own Model.

How long a turn may run

Settings → AI → Turn time limit → Stop a turn after sets how long Lucius may work on one reply or one Automation run before it’s stopped: 20 minutes (default), 30 minutes, 1 hour, 2 hours or No limit.

A stopped turn says so — “This turn hit its 20-minute time limit and was stopped.” — and nothing is left half-applied. Most turns finish in seconds; one that keeps hitting the limit is usually waiting on a slow model. Pick a faster model (for an Automation, set its own Model) or raise the limit. With No limit, a turn runs until it finishes or you press Stop. A single call to the Claude Code or Codex CLI is still stopped after 20 minutes, since a call that long means the CLI has hung.

Long conversations

A long conversation is compacted automatically: older messages are folded into a summary while recent ones stay word for word, so it keeps working and doesn’t get more expensive with every message. There’s nothing to set. This is separate from Memory, which carries facts between conversations.

See what it costs

Per reply: under each reply, the tokens, price and model that answered. Turn it off with Settings → Chat & Memory → Show cost per turn — hiding it changes nothing about what’s spent or recorded.
Per conversation: the running total next to the model name in the message box.
Per Automation: its Cost & usage section; sort the Automations list by Cost this month.
Everything: Settings → Cost & Performance. Pick a Period (Today, Yesterday, 7 days, 30 days, All time) to see AI spend — total, turns, tokens and cache hits, broken down By surface, By provider and By model, a cost trend, and a Projected monthly estimate at that period’s rate. Below it, Search & fetch shows estimated charges from web-search providers, and Model performance shows each model’s speed and reliability.

Prices are calculated from each provider’s published rates. With the Claude Code CLI on a Claude plan, the figure is what that usage would cost at list price — your plan covers it, and its usage limits apply instead of a bill.

To cut costs: turn off Follow-up suggestions in Settings → Chat & Memory (that stops a model call per reply), route background tasks to a cheaper model, and use Low effort for simple chats.

Troubleshooting

Key rejected — the API key is wrong or was revoked. Edit the provider and paste a new one.

Can’t connect — the endpoint isn’t reachable. For a local model, check it’s running and the Base URL is right (Ollama’s is http://localhost:11434/v1).

The connection test failed — the message under it says why. Report this problem opens the problem report with that error filled in, for you to review before sending.

“Choose a default model on another provider first” — you can’t switch off the provider your default model uses. Pick a different Default model, then switch it off.

”… can’t read images” — an image task (such as Image reading) is set to a text-only model. Choose a model with vision for it in Task models.

Connecting an AI provider — setup for each provider
Memory — has its own background model
Automations — each can use its own model