Voice callsExperimental
Experimental. Voice calls are off by default. Turn them on in Settings → Developer → Experimental → Voice calls. While the switch is off, nothing on this page appears in the app.
Press Call in Lucius’s chat and talk instead of typing. Lucius answers out loud, looks things up, and proposes changes exactly as he does in the chat — and the call shows up in that same conversation as it happens, so you can read along and approve cards on screen.
How a call works
A call has two voices working together:
| Who | What it does |
|---|---|
| The front voice | A fast model that keeps the conversation natural: it greets you, answers “are you still there?”, lets you interrupt, and says goodbye. It answers in about a second. It never sees your data and never states a number or claims something is done. |
| Lucius | Does every real thing — balances, spending, budgets, investments, changes — with the same tools, permissions and approval cards as the chat. His answers are read out word for word, so an amount is never re-phrased. |
The whole call is in the chat — the greeting, small talk the front voice answers itself, and every request with Lucius’s answer — so you can scroll back through it later, and Lucius knows what was said on the call next time you talk.
When you ask for something real, the front voice hands it to Lucius and says something short (“Sure, one sec”). Lucius’s question and answer appear in the chat. If a lookup is genuinely slow, you hear “the live prices lookup is taking longer than usual” — once, and only when it’s true. Ask “what are you doing?” and the front voice tells you what Lucius is working on right now.
Changes still need your approval. When Lucius proposes one, he says so, and the approval card appears in the chat on your screen. Tap Approve and he tells you on the call when it’s done. Your permission settings apply exactly as they do in the chat.
You only ever talk to Lucius: the front voice speaks as him, so you’ll hear “let me check”, never “let me ask Lucius”.
Say “bye” or “that’s all” and Lucius says goodbye and hangs up. Closing the tab ends the call too. If the line goes quiet for about 20 seconds he asks “are you still there?”, and after another 20 he says goodbye and hangs up — never while he’s still working on something for you or a change is waiting for your approval. Need a moment? Say “hold on” or “give me two minutes” and he waits quietly for as long as you asked (up to 10 minutes) before checking in.
Set it up
Voice calls run on your own LiveKit project. LiveKit carries the call’s audio and runs the speech-to-text, the voice and the front voice model, so there is no other account to create. Its free plan includes a monthly allowance.
-
Install the voice libraries on the machine Open Net Worth runs on, then restart the app:
Terminal window pip install -r requirements-voice.txt -
Create a LiveKit project at cloud.livekit.io and, in its Settings → Keys, create an API key. Keep the URL (
wss://…livekit.cloud), the API key and the API secret. -
In Open Net Worth, open Connectors → Voice calls and paste the three values. Save.
-
When the page says Ready for calls, press the phone button in Lucius’s chat — it sits where Send is while the message box is empty. Your browser asks for the microphone the first time.
During a call the message box becomes a call bar: Ringing…, then Connected · 0:42 (or Lucius is working… while he looks something up), with Mute and End call. Your draft is kept and comes back when you hang up. If a call can’t start or drops, a note at the top of the box says why, with a button to try again, open Voice calls settings or view the app’s logs; dismiss it with ×.
In the Open Net Worth desktop app, the first call asks macOS for the microphone (once — it remembers). The desktop app allows the microphone for the app’s own page only, and only audio; this needs desktop version with voice-call support or later.
The browser only allows the microphone on a secure address: localhost on your own computer
works; from another device the app must be reachable over HTTPS (see
Exposing beyond localhost).
Settings
| Setting | What it changes |
|---|---|
| Front voice | The model that holds the conversation. Gemini 2.5 Flash-Lite is the default: it was the fastest in testing and the only one that asked “go ahead with what?” instead of guessing. |
| Lucius on calls | The model Lucius answers with during a call. Unset, he uses the conversation’s own model. A faster one keeps the waits short. |
| Language | What you speak on calls: English, Vietnamese, Chinese, Japanese, Korean, Hindi, Indonesian, Spanish, French, German or Portuguese. Lucius hears it, his greeting and check-ins are in it, and his voice speaks it. |
| Speech to text | Which service turns your voice into text. Automatic lets LiveKit pick the best one for your language; AssemblyAI Universal (English) hears English only, so another language moves you to Automatic. |
| Voice and Voice id | Which service speaks, and which of its voices (blank = its default). |
The LiveKit key and secret are stored like every connector credential: never shown to Lucius, readable back only by you through the eye on the field. Pausing the connector stops taking calls.
What reaches whom
- The call’s audio, the front voice’s side of the conversation, and Lucius’s spoken answers cross your LiveKit project and the speech and model services it runs.
- Your book itself stays with Lucius, as it does in the chat. The front voice has no access to it.
- The conversation is saved like any chat with Lucius.
Troubleshooting
The Call button isn’t there. Voice calls are off: turn them on in Settings → Developer → Experimental. It also gives way to Send as soon as you type.
The button says “Set up voice calls”. Voice calls are on but not set up yet: pressing it opens the Voice calls settings right where you are.
“Set up Voice calls in Connectors first.” The connector has no LiveKit URL, key or secret yet, or it is paused.
“The voice worker isn’t running.” Open Connectors → Voice calls: it says why — usually the voice libraries aren’t installed yet (step 1 above), or the LiveKit details are wrong.
A call dropped, then voice calls came back on their own. The voice side of the app runs separately from the rest of it, so a crash there ends the call but never the app; it restarts itself within a few seconds. Call again.
Lucius takes a while to answer. The front voice answers at once; Lucius’s own answer takes as long as a chat reply. Pick a faster model under Lucius on calls, and close heavy programs — a very busy computer also slows how fast the call notices you’ve stopped talking.
Related pages
- Telegram — chat with Lucius from your phone
- Confirm cards — approving what Lucius proposes
- How Lucius’s permissions work