Frequently Asked Questions

Find answers to the most common questions about LIA

Getting Started

How to get started with LIA

What is LIA?
LIA (Learning Intelligent Assistant) is your personal intelligent assistant powered by artificial intelligence. LIA helps you daily with:

• Managing your emails: search, read, send, reply and forward
• Organizing your calendar: create, modify and view your events
• Managing your contacts: search, create and update
• Managing your tasks: create, track and complete your to-dos
• Exploring your files: search, read and organise your files
• Getting information: weather, Wikipedia, web search, places

LIA understands natural language: talk to it like you would a human assistant!
How do I start a conversation with LIA?
It's simple in 3 steps:

1. After signing in, open Chat from the top bar
2. Type your message in the input box at the bottom of the screen
3. Press Enter to send

💡 Tip: Be natural! LIA understands phrases like "Show me my emails from this week" or "What do I have scheduled tomorrow?"
Does LIA support multiple languages?
Yes! LIA supports 6 languages:

🇫🇷 French (Français)
🇬🇧 English
🇪🇸 Spanish (Español)
🇩🇪 German (Deutsch)
🇮🇹 Italian (Italiano)
🇨🇳 Simplified Chinese (简体中文)

To change language:
• Click the globe icon 🌐 in the title bar
• Or go to Settings > Appearance > Language

LIA will respond in the interface language.
What can I ask LIA?
LIA can help you in many areas. Here are concrete examples:
🌐 MULTI-DOMAIN
  • "What will be the best day for an outdoor walk this week and what places near me would be good for this walk (ideally near somewhere we could enjoy a coffee and some pastries)"
  • "Tell me the history of the Eiffel Tower and find 2 nearby hotels and 3 restaurants with a rating above 4.5 nearby"
  • "Show the last 5 emails from my wife"
  • "Send an email to my wife and son to explain that I'll be late. Write it in a way that they won't be upset and use a humorous tone to soften the blow."
  • "Schedule an appointment with my son at the doctor Friday from 3pm to 4pm"
  • "Reschedule today's 2pm meeting to 4pm and change the location to room 38"
  • "Send an email to the participants of my next meeting to let them know I'll be 30 minutes late and extend the meeting by 30 minutes"
  • "What lunches do I have planned this week?"
  • "What are my free slots on Friday?"
  • "Show my contacts who work at Cap Gemini"
  • "When should I leave to arrive at Bastille tomorrow at 9pm by subway?"
  • "When am I planning to visit my brother?"
  • "Give me the TV schedule for tonight around 9pm on the first 6 channels"
  • "Give me the weather for Paris, London and New York tomorrow"
  • "Create a reminder the day before at 7pm for my next 2 appointments"
  • "Find 2 restaurants near my brother's place and give me the cycling directions from my house"
  • "Give me the driving route and weather for my next 2 appointments"
  • "Find the 5 best restaurants within 600 meters around here"
  • "Find 2 hotels with a rating above 4 within 10 km of Paris"
  • ...
👥 CONTACTS
  • "show contact John Smith"
  • "find John Smith's phone number"
  • "what is John Smith's address?"
  • "search for Smith contacts"
  • "who is the contact with phone number 555-0123?"
  • "search my contacts for where the Smiths live"
  • "find the Smiths who live together"
  • "find the Smiths who live in California"
  • ...
  • "create a contact"
  • "change John Smith's phone to 555-0199"
  • "change the third contact's phone to 555-0199"
  • "update my brother's address to 2 Oak Street in Boston (memory)"
  • "delete contact John Smith"
  • "delete the second listed contact"
  • ...
📧 EMAILS
  • "what are my last 2 received emails?"
  • "my last 3 unread emails excluding spam"
  • "search my last 3 sent emails"
  • "search the last 3 received emails with the personal label"
  • "find the last 3 emails from my brother (memory)"
  • "show the last 3 important unread emails"
  • "search my brother's last 3 emails about the budget (memory)"
  • "list the last 3 emails sent last week"
  • "find the last received email and get the contact details of the sender"
  • ...
  • "send an email"
  • "send an email to my brother and John Smith to explain I'll be late. Write it so they won't be upset. Use a casual tone. (memory)"
  • "send a romantic poem by email to my wife (memory)"
  • "reply to the last received email that I won't be able to attend and apologize. Use a formal tone"
  • "forward the last received email to my personal inbox (memory)"
  • "forward the last sent email to John Smith"
  • "delete the last two received emails"
  • "delete the second listed email"
  • "delete unread emails that are commercial"
  • ...
📅 CALENDAR
  • "show my calendars"
  • "find my next 2 appointments in the work calendar"
  • "what do I have planned today"
  • "my meetings this week"
  • "show my appointments for next week"
  • "what do I have planned next week?"
  • "search my schedule for January 29"
  • "find appointments with John Smith"
  • "what are the upcoming events with my brother? (memory)"
  • "find appointments in Paris"
  • "what are my availabilities next week?"
  • "do I have a free slot for a meeting Thursday between 2pm and 3pm?"
  • "find the next doctor's appointment"
  • ...
  • "add an event friends night tomorrow from 9pm to midnight"
  • "schedule a lunch with John Smith Monday from noon to 2pm at The Bistro"
  • "create a weekly sync meeting for one hour tomorrow at 2pm with John Smith at Main Street"
  • "modify the weekly sync event scheduled Tuesday, reschedule to Thursday at 3pm"
  • "extend the weekly sync meeting by one hour"
  • "add [email protected] as a participant to the weekly sync meeting"
  • "move my weekly sync meeting tomorrow from 2pm to 4pm"
  • "change the location of the weekly sync meeting to room B"
  • "remove [email protected] from the weekly sync meeting"
  • "cancel the weekly sync meeting"
  • ...
✅ TASKS
  • "show my tasks"
  • "what do I have to do?"
  • "list my current tasks"
  • "list my completed tasks"
  • "list tasks with overdue deadline"
  • "list today's tasks"
  • "list tasks due before the end of this week"
  • ...
  • "add a task call the bank"
  • "create a task renew my subscription with deadline December 31"
  • "create a task prepare the presentation with deadline Friday"
  • "modify the task prepare the presentation, push the deadline by one day"
  • "mark the task prepare the presentation as completed"
  • "delete completed tasks"
  • "delete the task renew my subscription"
  • "I finished the first listed task"
  • "I've completed all tasks"
  • ...
📁 FILES
  • "list my files"
  • "show my folders"
  • "search for PDF files"
  • "my 5 most recent files"
  • ...
🔔 REMINDERS (with notification)
  • "remind me to walk the dog at 7pm"
  • "remind me to call my brother Friday at 6pm"
  • "remind me to take my car to the mechanic on Feb 13 at 8am"
  • ...
  • "delete all my reminders"
  • "delete the reminder to call John Smith"
  • "delete the second listed reminder"
  • ...
📍 GOOGLE PLACES
  • "where am I?"
  • "what address am I at"
  • "find 2 restaurants nearby"
  • "find 2 Chinese restaurants with a rating above 4"
  • "find pharmacies open now"
  • "what time does The Bistro open?"
  • "how far is the nearest open restaurant?"
  • "where is the nearest gas station?"
  • "show supermarkets within 1 km"
  • "list 5 Italian restaurants in downtown Manhattan"
  • "find 3 hotels near the Eiffel Tower"
  • "search for museums in central London"
  • ...
🗺️ GOOGLE ROUTES
  • "how do I get to Paris?"
  • "how do I get to my brother's place? (memory)"
  • "what's the route to John Smith's house?"
  • "how do I get to Paris by public transit?"
  • "give me directions from Paris to London"
  • "how long to get to Paris by bike?"
  • "get to Times Square on foot"
  • "when should I leave to arrive at Piccadilly at 9pm by public transit?"
  • "get to Lyon without tolls"
  • "what's the route Paris-Berlin without tolls or highways?"
  • "what's the travel time to Paris by car right now?"
  • ...
🌤️ WEATHER
  • "what will the weather be like today?"
  • "give me the hourly weather"
  • "search the weather for tomorrow"
  • "search the weather for the next 3 days"
  • "what's the weather this week?"
  • "what will be the best day for an outdoor walk?"
  • "will it rain this week?"
  • "when will it stop raining?"
  • "which day will be the least windy?"
  • "search the weather in Paris, London and New York tomorrow"
  • ...
🔍 PERPLEXITY
  • "search for the latest iPhone news"
  • "search online for tomorrow night's TV schedule for the first 3 channels"
  • "what happened in the last French national team soccer match?"
  • "make a comparison of the best electric cars"
  • "search for the Louvre museum opening hours"
  • "search the web for the price of a Paris-New York plane ticket"
  • "how old was Marie Curie when she won her Nobel Prize?"
  • ...
📚 WIKIPEDIA
  • "search Wikipedia for information about the Eiffel Tower"
  • "find the complete Wikipedia article about the Louvre"
  • "find info on Wikipedia to explain photosynthesis"
  • "tell me about the French Revolution"
  • "who is Marie Curie?"
  • "search Wikipedia articles related to space exploration"
  • "on Wikipedia what are the related topics to the Renaissance?"
  • ...
How does the approval system work?
With human review (Human-in-the-Loop), LIA shows a draft for operations requiring your approval. You check the proposed content and scope before confirming. Read operations and declared direct actions can proceed without a draft card.

🔒 Actions requiring approval:
• Send, reply to or forward an email
• Create, modify or delete an event
• Create, modify or delete a contact
• Create or complete a task

📋 How it works:
1. LIA prepares a draft of the action
2. You see a detailed preview of what will be done
3. You can:
• ✅ Approve: the action is executed
• ✏️ Edit: correct before validation
• ❌ Cancel: nothing is done

💡 Tip: Take time to review the preview, especially for emails!
Can LIA make mistakes?
Like any AI assistant, LIA may sometimes misinterpret a request. Here's how to get the best results:
✅ Best practices:
  • Be specific: "Send an email to [email protected]" rather than "Send an email to John"
  • Give context: "Find my emails from this week about the Alpha project"
  • Review previews before approving actions
⚠️ If there's an error:
  • Rephrase your request with more details
  • Use specific dates ("January 15") rather than relative ones ("last week")
  • For important emails, always review the preview
🛡️ Built-in security:
  • Actions follow the approval rules that apply to them. For MCP, explicit saved choices can allow execution without another question; review their behavior summaries in Settings.
How was LIA built?
Nearly 100% of LIA's code is written by an AI, under human direction: written engineering rules, blocking automated checks, a review cycle where the human decides, and recurring adversarial audits (latest assessment: 7.9/10 across 24 areas — full report).

The full story — method, trade-offs, operations and what remains to be done — is on the Story page, and the source code is open on GitHub.
Can I install LIA on my phone — as a PWA or a real app?
Yes, two ways. As a web app (PWA): open your LIA in the browser and use “Add to Home Screen” — it launches full-screen and, on iOS 16.4 and later, can receive web push. Without a connection it shows a simple offline page: your conversations and other personal data are never stored on the device for offline reading.

As a native app: install the LIA app from the store and type your server's address once at the first screen. You get the same interface plus what only a native app can do — Google sign-in in the system browser, native notifications (on Android straight from your server's own Firebase project; on iOS through a wake relay that carries no content, when your server is set up with one), and connectors that return to the app after authorization.

Your server stores the account data; the app connects to it, and configured external services receive the context needed for enabled features under their own policies.
My account says it is awaiting activation — what should I do?
Signing up creates your account, but an administrator has to activate it before you can use LIA. Until then you land on a page that explains where you stand, and the rest of the application stays closed.

• You will receive an email as soon as your account is activated. Check your spam folder if you do not see it after a while — it is the most common reason people think nothing happened.

• Nothing is lost in the meantime. Your account already exists; activation only opens the door.

• If several days pass with no email, contact the administrator of your organisation, who can activate the account directly.
Is there a mobile app for Android or iPhone?
Yes — one official app per store, and it works with any self-hosted LIA server. At first launch you type your server's address (HTTPS required); the app then loads your server's own interface, so every web improvement reaches it instantly with no store update. Google sign-in, notifications, and returning from a connected service (Gmail, calendar, MCP…) all work natively. If your server is unreachable, a built-in offline screen offers a retry and a way to switch servers. The web app and the PWA keep working exactly as before — the apps are an addition, never a replacement.
How do I sign in from the mobile app?
With the same account as on the web. Signing in with your e-mail and password happens right in the app.

🔐 With Google: Google refuses sign-in inside an app's built-in browser, so the app opens your phone's own browser, then brings you back automatically once you are signed in. The return link carries a single-use code that only your app can redeem: intercepted, it is worthless.

🔢 Your second step still applies: if you set up an authenticator app in Settings › Preferences › Security › Strong authentication, its code is asked here too — signing in with Google never skips it, in the app or on the web.
Does LIA work without a connection once installed?
No — LIA runs on your server, and what you install is a window onto it. Conversations, answers, actions and synchronisation all need the network.

📴 When the connection drops:
• installed from the browser, LIA shows its own offline page, in your language, with a Try again button instead of a browser error
• the mobile app shows its offline screen, with a retry and a way to choose a different server — handy if the address was mistyped

🔒 Nothing personal is kept for offline reading: at most the application's own files are stored on the device; your conversations and data stay on the server.

When you reconnect, the open conversation catches up with what happened in the meantime.
What can I do in LIA World?
Open 3D from the menu when your instance offers it. Explore the shared house and garden, meet other people, exchange private or public text and open your personal LIA conversation and native tools. Choose your character and private home palette. Keyboard, touch, zoom and the minimap help you move around. Lighting follows your account’s local time; available cached weather and optional sounds animate the scene. Movement itself calls no model; activities and translation keep their normal permissions, confirmations, costs and quotas. The LIA World section below explains each feature.

LIA World

A shared place to meet, with your own private LIA close at hand.

What is LIA World, and how do I enter?
LIA World is a shared isometric house and garden where places lead to LIA’s functions. Open 3D in the menu when your instance offers it. Choose your character, review the displayed LIA account name and accept sharing your name, appearance and coarse presence. Change your name in your LIA profile settings. The operator sets capacity; when full, a queue offers a place for you to confirm. Leaving releases your place and keeps you signed in to LIA.
How do I move, look around and sit down?
Focus the scene and use arrows, WASD or ZQSD, or click or tap reachable ground. A mobile joystick is available. Drag to explore the camera; use the wheel, pinch or zoom buttons to change scale. The minimap moves the view; recenter follows your character again. Nearby authored seats and resting places offer their available poses. Help explains the controls and lets you adjust graphics quality.
What can other people see about me?
People in the shared scene see your LIA name, chosen appearance and public movements. Your accepted connections can see a coarse status: in the world, at home, away or offline. Your private home interior, conversation with LIA, memories, documents and personal panels stay outside shared snapshots. Blocking filters visibility and correspondence. Instance operators may have technical access; text sent for translation reaches the configured AI provider under its policies.
What can I personalise in my home and character?
Choose your character’s skin tone, hair, clothing, colours and accessories at entry or through My character, or start from one of ten editable appearances. Your saved appearance is visible in shared places. Your home is private: preview its colour palette and save it, or follow your application theme. Saved choices return on your next visit; a concurrent change asks you to reload before saving. This version offers bounded palette choices; home visits and shared furniture editing are unavailable.
Can I use LIA’s normal functions without leaving the world?
Yes. The places and function directory open the existing LIA panels privately: knowledge spaces, memories, files and kept answers, meetings, workboard, relations, reminders, routines, connectors and extensions, according to your account’s capabilities. Your personal conversation uses the normal workspace, including available attachments and voice controls. Confirmations, permissions, quotas and costs still apply. Administration stays in the ordinary settings. You can return to the usual application at any time.
How do I talk to LIA, and can I speak to other people?
Call LIA or select the character to open your personal conversation, journals, notifications or psyche. LIA can patrol or follow you; walking the character itself calls no AI model. Dictation, spoken answers and Live sessions use LIA’s existing voice features when available and configured. They connect you to LIA. Communication between human participants in this version uses text; direct human audio is unavailable.
How do private and public messages work?
Open Conversations and choose a person visible in a shared place for a direct message, or the public conversation for a message to its eligible audience present at sending. A public message stays with that audience; later arrivals do not receive its history. Return from home to a shared place before sending. A quick bar also writes to the nearest reachable person close to you, otherwise to everyone present: it shows the destination, lets you change it and keeps it while you write or a message is sending. On a phone, Chat opens it. Original and translated text can be viewed, with pending, failed or uncertain delivery shown explicitly. These temporary messages expire after the duration shown in the panel and are not added to LIA’s memory.
Who pays for message translation?
Sending asks for consent to any translation charge under your LIA quotas. Same-language delivery calls no model; a translation uses the instance’s dedicated translator and shows its available token and cost receipt. A public send groups the required target languages into one translation request. A later translation you explicitly request uses your own quotas. Failure can still consume paid work; an uncertain request offers a status check rather than automatic resending.
What are the phone, weather and sound controls for?
The compact device panel gives private access to your SMS, reminders, notifications and available calling or radio controls. Proactive notifications can also be offered by LIA without exposing their content to the scene. The weather panel reads your available cached weather; day/night lighting follows your account’s local time, anchored to the server clock. The scene makes no extra weather-provider request. Optional ambience and message sounds have their own controls. Sound and animation do not create a shared microphone or publish your personal data.
What if the connection drops, another tab takes control or my device struggles?
The server confirms movement and collisions. Losing authority pauses input while the world reconnects; an expired place can require joining again. Another tab must explicitly take control. Changing account closes the old private workspace and asks you to reload with the current account. Help offers graphics quality and performance information; reduced motion limits decorative animation. Keyboard focus can leave the scene normally, and touch controls support phones. Observed performance depends on your device and connection.
How does an administrator make World available?
The operator enables World in server configuration, then the administrator permits it in Settings › Administration › Platform capabilities. The operator also sets capacity and the server limits; the capability switch cannot override a disabled deployment. World reuses native account sessions, Redis-backed presence and saved account preferences. Personal panels keep their own capability gates, while message translation and LIA's written remarks each use their own model slot and the existing usage limits. No administrator panel is exposed inside the shared scene.
How do I discover places and bring the scene to life?
The Functions menu lists places and tools available to your account, including SMS in the atrium. Route, Gestures and Photo sit with the widgets. The route is enabled by default and can be switched off. Click interactive scenery: your character approaches before the reaction and sound, when audio is enabled. Lamps toggle and screens offer five displays. Interactions, seasons and local effects call no AI model. Your private tools stay private.
How do the arcade cabinets work?
Each cabinet has its own game: tic-tac-toe, four-in-a-row or Reversi. Choose the computer when alone, or wait for another person and explicitly accept a request. A new human match starts with a fresh board. The server confirms legal moves and results. The board shows where you may play and previews your piece before you place it; in four-in-a-row, click anywhere in a column. Reversi keeps each side’s count, and the winning line is highlighted. After a connection interruption, explicitly rejoin; leaving withdraws your participation. Playing calls no language model.
How do I take and share a private photo?
Photo prepares a local landscape or portrait JPEG with your character, LIA and the scenery. Framing respects your selected zoom; the format may crop the edges. The signature LIA World - https://lia.jeyswork.com is always present. Other visitors, names, bubbles and private panels are excluded. Preview it, then save or choose Share when offered. The preview uploads nothing. Opening a system chooser does not confirm publication by another application.
How do gestures work?
Open Gestures with the widgets, or use the buttons beside the joystick on a phone. Wave, thank, applaud, laugh, ask or celebrate: the server confirms the gesture and shows it above your avatar to the people who can see you, for a few seconds. Gestures are limited to a calm pace; when one is refused, a short message tells you why. They are not shown inside your private home and call no AI model.
What happens when I sit at a table?
When you sit at an ordinary table, LIA asks whether you want to open your full dashboard: briefing, radio, results and consumption. Choose Stay seated to decline; nothing is loaded or billed. Accept, and the dashboard opens privately with its usual costs and confirmations; you can offer it to yourself again without standing up. Familiar objects also show your own signals — a reminder that fired, a new message, a ticket waiting for you — visible only to you.
What does LIA do when I ask for nothing?
LIA lives a life of its own in the world: it strolls from place to place — the fountain, the library, the boardwalk on the pond, the kitchen garden, the telescope after dark — stops, sits on a bench or lies down in the grass, looks around and sometimes plays with what is there. Now and then a short line appears above LIA about the place, the hour or the weather; it fades by itself. As soon as you give LIA real work, the stroll stops on the spot and resumes when the work is done. Follow me calls LIA to your side. The strolls are set by the scene and call no AI model. The lines can be written by a dedicated AI model when your instance offers it: LIA then says them with the personality you chose, a few at a time, for its next stops only. The model receives that personality, your language, the places, and the hour, weather and season the world already shows — none of your data. Those few tokens count in your usage, under your usual limits; beyond them, or when the model is off or unavailable, LIA says lines written in advance, which cost nothing. None of this is stored or shared. Around it the garden is alive too: ducks, a frog, fireflies at dusk… and a garden gnome that never stays put.
Do things in the world react when I touch them?
Yes, almost everything answers. Click or tap an object: your avatar walks up to it, then the object does its turn. The pink flamingo spins on its leg, the mole comes out of another molehill to stick its tongue out at you, the bees chase you for a few seconds, the garden gnome plays dead, a jack-in-the-box springs out of the theatre trunk, a spotlight follows you… and at night, something comes for the flamingo. Trees can be shaken, bushes have eyes, and the grass bends as you and LIA walk through it. LIA sets these turns off too on its walks, and its face reacts: it blinks, its eyes follow you, its visor shows hearts, stars, spirals or chevrons depending on its mood, and its mouth moves while one of its remarks is on screen. Play a prank in front of LIA and it may have something to say about it. With the World's sound on, each turn makes its own noise too: the flamingo squeaks, the mole blows a raspberry, the bees buzz, the crow caws… All of this is purely decorative: it plays on your screen only, calls no AI model, costs nothing and records nothing. If your device asks for reduced motion, objects stay still.

Connectors

Integrations and external connections

What is a connector?
A connector is a secure gateway between LIA and your external services:
🔌 Types of connectors:
  • Google Services (OAuth):
  • 📧 Gmail: read, send, reply to emails
  • 📅 Calendar: manage your events
  • 👥 Contacts: search and create contacts
  • 📁 Drive: explore your files
  • ✅ Tasks: manage your tasks
Apple iCloud services (app-specific password):
  • 📮 Apple Mail: read, send, reply to emails (IMAP/SMTP)
  • 🗓️ Apple Calendar: manage your iCloud events (CalDAV)
  • 📇 Apple Contacts: search and create iCloud contacts (CardDAV)
Microsoft 365 services (OAuth):
  • 📧 Outlook: read, send, reply to emails
  • 📅 Microsoft Calendar: manage your events
  • 👥 Microsoft Contacts: search and manage contacts
  • 📁 OneDrive: explore and organise your files
  • ✅ Microsoft To Do: manage tasks and lists
External services (API key):
  • 🌤️ OpenWeatherMap: weather and forecasts
  • 🔍 Perplexity: AI web search
  • 🦁 Brave Search: web and news enrichment
Smart Home (hybrid auth):
  • 💡 Philips Hue: control lights, rooms, and scenes
Telephony (bring-your-own):
  • 📞 ElevenLabs Telephony: LIA places outbound phone calls on your behalf — see the dedicated Phone Calls section
Always active, nothing to configure (provided by your instance):
  • 📚 Wikipedia: encyclopedic search
  • 🌐 Page browsing: open and read web pages
  • 📍 Google Places: search for locations
  • 🗺️ Google Routes: directions and routes
  • 🌤️ Google Weather: weather and forecasts
  • 🌿 Google Environment: air quality and pollen
🔒 Security:
  • Each connector has limited permissions. LIA only accesses necessary data.
⚡ Mutual exclusivity:
  • For email, calendar, and contacts, only one provider can be active at a time (Google OR Apple OR Microsoft, not both simultaneously). Tasks are available via Google Tasks or Microsoft To Do, files via Google Drive or OneDrive (one at a time).
How do I connect Google services?
Google services use OAuth 2.0 (secure connection):
Steps:
  • 1. Go to Settings > Connectors
  • 2. Find the desired service (Gmail, Calendar, etc.)
  • 3. Click Connect
  • 4. A Google window opens
  • 5. Select your Google account
  • 6. Accept the requested permissions
  • 7. You're redirected to LIA - the connector is active!
🔐 Permissions:
  • LIA only requests necessary permissions. For example, for Gmail: reading and sending emails, but not permanent deletion.
💡 Tip:
  • You can revoke access at any time from LIA settings or from your Google account.
How do I connect external services (API key)?
Some services require a personal API key:
🌤️ OpenWeatherMap (free):
  • 1. Create an account at openweathermap.org
  • 2. Get your free API key
  • 3. In LIA: Settings > Connectors > OpenWeatherMap
  • 4. Paste your key and click Activate
🔍 Perplexity (paid):
  • 1. Create an account at perplexity.ai
  • 2. Subscribe to an API plan
  • 3. Get your API key
  • 4. Activate in LIA as above
📚 Without a key:
  • Wikipedia, page browsing and the Google services your instance provides (Places, Weather, Environment) are always active: there is nothing to activate. When your administrator runs the instance's web search (SearXNG), it finds your web links too, unless you configured your own Brave key.
🔒 Key security:
  • Your API keys are stored securely and encrypted.
How do I disconnect a service?
To revoke LIA's access to a service:
Disconnection from LIA:
  • 1. Go to Settings > Connectors
  • 2. Find the service to disconnect
  • 3. Click Disconnect
  • 4. Confirm disconnection
⚠️ Consequences:
  • LIA will no longer be able to access this service
  • Your data remains intact with the provider (Google, etc.)
  • You can reconnect at any time
🔐 Double security (Google):
Why is a connector showing an error?
Several reasons can explain a connector error:
🔁 Renewal is automatic:
  • LIA renews your access by itself, once, even when several features need it at the same moment, and marks a service in error only after the provider confirmed its refusal — so a Reconnect request means your authorization really needs you.
🔄 Expired token:
  • Solution: Click Reconnect to refresh the authorization.
🚫 Permissions revoked:
  • If you revoked access from Google, reconnect.
🔑 Invalid API key:
  • For API key services, verify your key is correct and active.
⏱️ Quota exceeded:
  • Some services have usage limits. Wait or upgrade to a higher plan.
🌐 Network issue:
  • Check your internet connection and try again.
💡 General solution:
  • Disconnect then reconnect the service. This solves most problems.
Can I configure preferences per connector?
Yes! Some connectors have customizable preferences:
📅 Google Calendar:
  • Default calendar: where to create new events
  • Example: "Work", "Personal", "Family"
  • LIA will use this calendar unless specified otherwise
  • To know when you are busy (free slots, your day, the meetings with a person), LIA reads every calendar you own, not this one alone
✅ Google Tasks:
  • Default task list: where to create new tasks
  • Example: "My main list", "Projects", "Shopping"
📍 Location:
  • Home address: your default address for "near me" searches
  • Automatic geolocation: send your position for accurate results
  • Location memory (opt-in): remember your last position so LIA keeps answering from where you are when the live position is unavailable — always announced with its age, never presented as current, encrypted and erased the moment you switch it off
  • Configure all three in Settings > Preferences > Location — this data lives on your account, not on a connector
🌐 Browser Geolocation:
  • In the Location section (Settings > Preferences), enable the geolocation toggle
  • Your browser will ask for permission the first time
  • Position used for local weather and "near me" searches
  • On mobile, the position refreshes silently when you return to the app; if the system dropped the permission, a chat banner offers to re-enable it
To configure:
  • 1. Go to Settings (⚙️ icon in the menu)
  • 2. Open the Connectors section
  • 3. Find the relevant connector (Calendar or Tasks)
  • 4. Configure the specific preferences for that connector
💡 Tip:
  • These preferences save you time by not having to specify the calendar, list, or city each time.
Which Google services can I connect?
LIA integrates with 7 Google services:
📧 Gmail
  • Search all your emails
  • Read message content
  • Send, reply, forward
  • View attachments
📅 Google Calendar
  • View your schedule
  • Create events
  • Modify or delete
  • Invite participants
👥 Google Contacts
  • Search by name, email, company
  • Create new contacts
  • Update information
📁 Google Drive
  • Search your files
  • Read document content
  • Browse folders
  • Organise them: folders, rename, move, copy, share
✅ Google Tasks
  • List your tasks
  • Create new tasks
  • Mark as completed
📍 Google Places
  • Search for locations
  • Get addresses and hours
  • View reviews
🗺️ Google Routes
  • Calculate routes
  • Compare transport modes
  • Estimate travel times
ℹ️ Places and Routes need no connection:
  • they run on your instance's Google key and are always active, like Google Weather and Google Environment.
Are connectors secure?
Security is our top priority:
🔐 Google services — OAuth 2.1:
  • LIA never sees your Google passwords
  • You connect directly with Google
  • Tokens revocable at any time
🍎 Apple services — App-specific password:
  • Uses an application-specific password (not your Apple password)
  • Encrypted before database storage
  • Revocable from appleid.apple.com
  • Apple two-factor authentication is required
🪟 Microsoft 365 — OAuth 2.1:
  • LIA never sees your Microsoft passwords
  • You connect directly with Microsoft
  • Tokens revocable at any time from your Microsoft account
  • Limited permissions (Microsoft Graph API)
🔒 Minimal permissions:
  • Each connector only requests necessary permissions
  • No excessive access to your data
🗄️ Secure storage:
  • OAuth tokens and app-specific passwords encrypted in database
  • API keys stored securely
  • No plaintext storage
🔄 Automatic refresh:
  • Google and Microsoft tokens expire regularly and are renewed automatically
  • Apple app-specific passwords don't expire (unless Apple ID password changes)
📋 Audit and logs:
  • All actions are tracked
  • Usage history available
How do I connect Apple iCloud services?
Apple iCloud services use an app-specific password:
Prerequisites:
  • An Apple account with two-factor authentication enabled
  • Generate an app-specific password from appleid.apple.com
Steps:
  • 1. Go to Settings > Connectors
  • 2. In the Apple iCloud section, click Connect
  • 3. Enter your Apple ID (email)
  • 4. Enter the app-specific password (format xxxx-xxxx-xxxx-xxxx)
  • 5. Select the services to activate (Email, Calendar, Contacts)
  • 6. Click Test connection to verify
  • 7. Click Activate
⚠️ Mutual exclusivity:
  • Activating an Apple service automatically deactivates its Google equivalent (and vice versa). For example, activating Apple Mail will deactivate Gmail.
🔐 Security:
  • Your app-specific password is encrypted and never exposed. You can revoke it at any time from your Apple account.
What Apple iCloud services can I connect?
LIA integrates with 3 Apple iCloud services:
📮 Apple Mail (IMAP/SMTP)
  • Search your iCloud emails
  • Read message content
  • Send, reply, forward
  • View attachments
  • Navigate by IMAP folders
🗓️ Apple Calendar (CalDAV)
  • View your iCloud calendar
  • Create events
  • Modify or delete
  • Invite participants
📇 Apple Contacts (CardDAV)
  • Search by name, email, company
  • Create new contacts
  • Update information
❌ Not available:
  • Apple Reminders: Apple doesn't provide a standard API for accessing Reminders from a server. Tasks remain available via Google Tasks or Microsoft To Do.
💡 Same experience:
  • Commands are identical whether you use Google or Apple. Just say "my emails", "my calendar", or "my contacts" — LIA automatically uses the active provider.
What is connector mutual exclusivity?
For email, calendar, contacts, and tasks services, LIA applies a mutual exclusivity rule:
⚡ Principle:
  • Only one provider can be active per category:
  • Email: Gmail OR Apple Mail OR Outlook (not multiple)
  • Calendar: Google Calendar OR Apple Calendar OR Microsoft Calendar (not multiple)
  • Contacts: Google Contacts OR Apple Contacts OR Microsoft Contacts (not multiple)
  • Tasks: Google Tasks OR Microsoft To Do (not both)
  • Files: Google Drive OR OneDrive (not both)
🔄 Switching providers:
  • If you activate a new provider while another is active for the same category:
  • 1. The previous service becomes inactive (not deleted)
  • 2. The new service becomes active
  • 3. Your data remains intact with each provider
  • 4. You can switch back at any time
✅ Unaffected services:
  • Places, Routes and Philips Hue have no equivalent in other providers and remain always available.
💡 Tip:
  • You don't need to choose the same provider for everything. For example, you can use Outlook for emails, Google Calendar for your calendar, and Microsoft To Do for tasks.
How do I connect Microsoft 365 services?
Microsoft 365 services use OAuth 2.1 (secure connection via Microsoft Identity Platform):
Steps:
  • 1. Go to Settings > Connectors
  • 2. In the Microsoft 365 section, click Connect All or select an individual service
  • 3. A Microsoft login window opens
  • 4. Sign in with your Microsoft account
  • 5. Accept the requested permissions
  • 6. You're redirected back to LIA — the connector is active!
🔐 Permissions:
  • LIA only requests necessary permissions via the Microsoft Graph API. For example, for Outlook: reading, sorting and sending emails, and your automatic replies and rules. Deleting for good is yours alone, in Tools, after a question.
⚠️ Mutual exclusivity:
  • Activating a Microsoft service automatically deactivates its Google or Apple equivalent for the same category (and vice versa).
💡 Tip:
What Microsoft 365 services can I connect?
LIA integrates with 5 Microsoft 365 services:
📧 Outlook (Microsoft Graph)
  • Search your emails
  • Read message content
  • Send, reply, forward
  • View attachments
  • Automatic replies and rules
📅 Microsoft Calendar
  • View your schedule
  • Create events
  • Modify or delete
  • Invite participants
👥 Microsoft Contacts
  • Search by name, email, company
  • Create new contacts
  • Update information
📁 OneDrive
  • Search and browse your files
  • Read document content
  • Organise them: folders, rename, move, copy, share
✅ Microsoft To Do
  • List your tasks
  • Create new tasks
  • Mark as completed
  • Manage task lists
💡 Same experience:
  • Commands are identical regardless of provider. Just say "my emails", "my calendar", "my contacts", or "my tasks" — LIA automatically uses the active provider.
How do I connect Philips Hue?
Philips Hue uses hybrid authentication (local press-link or remote OAuth2):
🏠 Local mode (recommended — same network):
  • 1. Go to Settings > Smart Home > Philips Hue
  • 2. Click Local connection
  • 3. Click Search for bridges
  • 4. Select your bridge
  • 5. Press the physical button on top of your Hue Bridge
  • 6. Click Pair within 30 seconds
  • 7. LIA validates connectivity and activates the connector
🌐 Remote mode (different network):
  • 1. Configure Hue Remote API credentials on the server
  • 2. Click Remote connection
  • 3. Authorize LIA via Philips OAuth2
💡 What you can do:
  • "Turn on the living room lights"
  • "Dim the bedroom to 30%"
  • "Set the kitchen light to blue"
  • "Set the Downstairs zone to blue"
  • "Activate the movie scene"
  • "What lights are currently on?"
  • "Turn off everything"
⚡ 6 tools available:
  • List lights, control a light (on/off, brightness, color), list rooms, control a room, list scenes, activate a scene.
Which connectors are always active?
The ones that ask nothing of you: Wikipedia, page browsing, Google Places, Google Weather and Google Environment (air quality and pollen) — no account to link, no key to paste. Your instance provides them to every account from the first conversation, so they do not appear in Settings > Connectors and there is nothing to switch on or off.

If one is missing, the instance is withholding it: its administrator switched it off for everyone, it has no platform Google key, or page browsing is disabled on it. For the weather, a personal OpenWeatherMap key you add in the settings takes precedence over Google Weather.
Can I connect several Google or Microsoft services with one authorization?
Yes. In Settings > Connectors, choose the Google or Microsoft services you want to connect or reconnect together, then sign in once to that provider account. LIA checks the account identity before sharing the authorization between those services.

You can disconnect one service without disconnecting the others. Existing connections keep working as before until you choose this new path. Your administrator must register the grouped callback address with the provider for the one-step flow to be available.

🔔 From the alert: when at least two services of one provider have expired, the connection alert shown in the app offers Reconnect all — the same path, in a single authorization, without going through the settings.
What do the email and appointment badges mean?
When the administrator enables the corresponding JEV integrations, emails can show a work, personal, spam or Phishing/scam category, with one priority: normal, urgent or critical. Normal priority stays visually quiet. An email that explicitly asks you for a reply or an action can also show Reply expected; spam and phishing never do. Appointments can show work or personal. An uncertain assessment shows no badge. The assistant receives the same accepted labels. These are advisory assessments of the visible content, not sender authentication or proof of fraud; no badge does not mean safe. The phishing badge is red, with a brief attention effect disabled under reduced motion.
Do I need a search key for LIA to search the web?
No, if your instance runs its own web search (SearXNG): LIA then finds there the web links of your searches, of the enrichment of its answers, of your interests and of the Tools. If you configured your own Brave key, it keeps priority. With neither, LIA only searches Wikipedia. Only the administrator can turn the instance's search on or off, in the connectors administration. Web results are not filtered, whether they come from your Brave key or from the instance's search.

Tools

Your mail, calendar, tasks, contacts, files and lights, by your own hand with no model, and with LIA in one click when you ask.

What is the « Tools » section?
Tools, in the header, opens your connected services directly so you can consult them and act on them yourself, as you would in your provider's own app. What you do by hand involves no model: it costs no token and nothing in euros, answers as fast as your provider does, and does exactly what you click. A row of tools above the open one switches between them in one tap — on a phone it scrolls sideways — and the section opens on the first tool that is ready. On a large screen, drag the line between two panes to widen one, or double-click it to give the pane its width back; this device remembers it. Refresh reads a tool again, and coming back to the page after a while does it on its own. The open tool's pill says whose service it is, and on a computer ? lists its keyboard shortcuts, which you can switch off there. Ask LIA, on a message, an event, a task, a contact or a file, opens the chat on it, ready for your words: nothing leaves until you send. Ctrl+K, or the magnifier beside the row of tools, searches your mail, calendars, contacts and files at once: each tool shows its first results, and says if it could not answer.
Why is a tool listed under « To connect »?
A tool works once its service is connected in Settings › Connectors: mail, calendar, contacts and tasks use the provider active for that category (Google, Microsoft or Apple — Apple has no tasks), files use the drive you connected, Google Drive or OneDrive, and lights your Hue bridge. Until then the tool waits under To connect, with a link to the connectors. A tool your administrator switched off is not shown at all.
Does LIA see or remember what I do in Tools?
What you do by hand is your act, not LIA’s: it adds no entry to LIA’s registers, memory, habits, journal or AI usage costs. This does not switch off features you enabled separately: the chat, routines, knowledge synchronisation and opted-in monitoring retain their own access and permissions. What you ask of LIA from these screens is LIA’s act: see “What can LIA do from the Tools?”.
What can I do with my mail and my calendar?
Mail: browse your folders (and labels with Gmail), search, read a message as text or in its original layout — shown in an isolated frame where nothing can run, its links opening in a new tab and its remote images waiting for your click, which tells the sender you opened it — mark read or unread, trash, label (Gmail), write, reply, forward with your own files attached, keep drafts, replies and forwards included, and with Gmail or Outlook set your vacation responder and filters.

Calendar: Agenda, Day, Week or Month, each calendar shown or hidden, events created, changed and deleted — all-day ones and video calls where your provider supports them. In a repeating series, a change reaches that occurrence alone unless you choose the whole series; with iCloud it always reaches the series, and the screen says so. In a message's recipients and an event's guests, typing a name suggests your contacts' addresses. The addresses in a message's text and in your notes open with a click, a filter can be deleted here, and a new event goes to the default calendar of your connector's preferences. A message's row has its own actions, Shift ticks a whole range, and on a computer the next message opens after one you trash. On a large screen the mailbox opens on your newest message, which stays unread until you open it yourself. The message you write stays beside the mail, which you can go on reading, and Send waits a few seconds so you can take it back. A message can become a task or an event, and its sender a contact. In the Day and Week views a line shows the time, a click on an empty slot starts an event there, and on a computer an event moves by dragging it and lengthens by its lower edge. If it has guests, the editor opens on the new time and nothing leaves until you save. Agenda lists the next thirty days, a video call opens from the event itself, an event can be duplicated, and on a phone Month shows the days with their dots, the picked day's events below. The images a message carries inside itself show on their own; only those it fetches from elsewhere wait for your click. An attachment opens in a window of the page, and a message's text prints from its menu. When a mailing list says how to leave it, Unsubscribe stands under the sender: one click after a question where the list promises it, else its own page or a message to send.

Sorting your mail: archive, move to a folder (from the menu, or by dragging a row onto the folder pane on a computer), star, report as spam or not, restore from the trash; Undo on the toast brings it back. Outlook and iCloud also delete for good and empty the trash, after a confirmation, and manage your folders; Gmail deletes what stays in its trash on its own after 30 days. The inbox shows its unread count on the Mail pill, and each folder its own.

Your calendar: answer an invitation from the event (accept, maybe, decline: the organiser is told), search your events, add a call to an event that has none, set a reminder, show the time as free, keep an event private, choose its time zone, send it to another calendar (on Outlook and iCloud, an event with guests or of a series stays where it is), create, rename and delete your own calendars, and see your dated tasks on their day.

Writing: attach a file from your device with the button or by dropping it on the message, on a reply and a forward too. A reply shows whom it reaches, and you can change its recipients and copies. A new message saves itself when you stop writing; a reply or a forward is kept with Draft, in its conversation, to finish from your drafts. A draft keeps its files. Formatting adds bold, italic, lists and links; what you paste arrives as plain words. Your signature, in the mailbox settings, closes your messages and those LIA prepares for you, where you read it on the card. With Gmail, choose the address a message leaves from when your mailbox has several. A message shows its conversation above its text, and the list can group each conversation into one line. An event can repeat, and a change or a deletion can reach the whole series. An Outlook connection made before these settings asks to be reconnected once for them.
And my tasks, contacts, files and lights?
Tasks: your lists and their tasks, ticked done (undone at once if your provider refuses), created, edited, deleted.

Contacts: sorted by last name then first name, each shown as first name then last name; browse, search, filter by group, find the « other contacts » met by mail (Google); every address, phone number and postal address is kept and edited one by one, and See the relationship opens the Relations page when the names match exactly.

Files, on Google Drive or OneDrive: your drive and its folders, a search over all of it, a file's preview and details, the trash with Undo. Create a folder, send files with the button or by dropping them on the list, rename, move, copy, share with someone or by a link, send a file by mail, and keep a mail's attachment on the drive. A Google document downloads as PDF or Word, a sheet as Excel, a presentation as PDF or PowerPoint. Google Drive adds its views (recent, starred, shared with you, shared drives), stars, and its trash, restored or emptied for good by hand after a question; OneDrive lists none of them, and SharePoint libraries are not listed. In the chat, LIA organises your files the same way and shows a card before it shares one.

Lights: rooms, lamps and scenes, with only the controls each lamp supports. Rooms and lamps come in alphabetical order, and a lamp's temperature and colour wait in a section you open. The tasks screen opens on All tasks; the default list of your connector's preferences is only where a new task goes when you name no list in the chat. An « other contact » opens read-only, and a scene no room or zone holds stays within reach under them. Tasks save as you leave a field, and today, tomorrow or next week are one tap away. A contact's addresses and numbers copy in one tap, a postal address opens on a map, and an e-mail address opens a new message. Files show what they are, sort by name, date or size, go to the trash several at a time, and their link copies. All off switches off everything lit, and on a phone a room's scenes are on its card. Once nothing is lit, the button becomes All on. Today, Overdue and Coming up gather the tasks of all your lists, All shows the three together, and All tasks, at the head of your lists, holds every task still to do, dated or not; each one names its list in a badge and opens beside the view, without leaving it; you can rename a list, delete its completed tasks, select several tasks at once and send a task to another list. With Microsoft To Do a task also keeps steps, an importance, a reminder and a repetition. A contact also holds a job, a website, a birthday and a photo; groups work with the three providers, and a vCard file exports or imports your address book. Zones sit after the rooms. A room or a zone takes a temperature or a colour as a whole, a lamp your bridge cannot reach says so, and a scene whose palette can play has a button that sets it in motion.
What if I click twice, or the connection drops while sending?
The screen guards repeated submissions to reduce duplicate sends and creates. When the earlier submission is still recorded, a second click is told “already submitted” instead of sending again. If the answer was lost, the outcome is shown as unknown: check Sent or the relevant service before submitting again. With the moment to take a message back, Undo stops it before it leaves; leaving the tool meanwhile sends it, since you pressed Send.
Can I read a file or an attachment without saving it, and change a text?
Yes. A file’s detail and a message’s attachments open in the page, on every device: a PDF whole, page after page, a Word file laid out as its author wrote it, a Markdown formatted, a spreadsheet or a CSV as a table, a sheet at a time, an image, a text. With Google Drive a Google document, and with OneDrive a Word file or a slide deck, arrive as a PDF the provider renders. A slide deck attached to a mail or kept on Google Drive reads as text. A long text says it stops before its end, and Enlarge gives the document a window. If a reading fails, Try again starts it over. A plain text file can be changed with Edit the text: if it changed elsewhere meanwhile, nothing is written and your words stay. An Office file or a Google document is changed where it lives, behind Open. LIA reads, adds to and rewrites a text file too, after a card showing the whole new text.
What does a contact’s detail show of our exchanges?
Your latest exchanges with them: the five latest messages they sent you, the five latest you sent them, the five latest meetings they organised with you invited, and the five latest you organised with them invited. A line opens in its tool. The calendar is read back over the last two years, in every calendar you own, and the screen says so. A meeting someone else organised answers neither list, and one that names no organiser stands in a list of its own. If your mailbox or your calendar could not be read, it says so rather than showing nothing.

With their numbers come the five latest calls LIA placed to them, each saying when, how it went and what it was for, and the five latest texts they sent you: a contact with numbers alone has exchanges too. A call opens in your calls list, a text on the SMS page. These lists appear once LIA has calls or texts to show, and never show the number itself.
What can LIA do from the Tools?
The yellow stars mark where LIA can help. The ones that end every bar open a list, Ask LIA last. There LIA offers what it can do on what you are looking at: sum up or translate a message, sum up a conversation, propose a reply, make an event or a task of it, read it aloud, sum up a file or an attachment, describe an image, complete a contact from their messages and, with Microsoft To Do, break a task into steps. Where you write, the same icon writes from an instruction, rephrases, translates, proposes a subject and lets you dictate. One sentence is enough to create an event or a task, search your mail or fill a filter, and a photo of a business card fills a new contact. So does a sentence, typed or said: a value is kept only if the sentence says it, a number by its digits. On a contact, Call asks LIA to phone one of their numbers once your telephony is connected: the chat opens on the words, and LIA asks before it dials. Write to … with LIA opens a new message ready for your instruction, and their notes have a writing helper too. Select words in a message's text or in a file shown in its panel: a bubble beside them explains, translates, simplifies or answers your question about them, without leaving what you read, and in a message it also makes a task or an event of them. A message shown in its original layout turns to its text with Act on a passage. On an event, Write to the guests with LIA opens a message to each of them; on a file, Send by mail with LIA attaches it to a message written from your instruction; on a task, Block a slot in the calendar opens a new event on its title and its day, which you save or not. In the list of messages, Sort puts the important ones first, groups by kind or keeps only the page's priority messages. A task or an event LIA reads in a message opens in its full form with everything the message says: who asks and what, the day, the place or the video link, the list it belongs to and, when the message sets up a meeting with you, its sender as a guest: saving sends them the invitation. A task found in a summary is read again for that action alone, and an instruction can be said rather than typed.

Nothing leaves without your click and nothing is written at your provider: LIA proposes, you read, you finish by hand. Each assist uses a model or a voice, only when you ask: what it costs shows beside the answer and counts in your usage, and what it read of your services stands in your Consultations, like any read of LIA’s. A value proposed for a contact comes with the message that says it, and is only proposed if that message really does. What you do by hand stays free of any model and any cost.

The same lists also offer Remind me, Ticket for LIA (a ticket on the board), Keep in a space and, on a message, Tell me when they reply. They end on a ready question for the chat, such as Prepare this meeting or Take stock: you read it, and nothing leaves until you send it. On an event, See the minutes opens the recorded meeting, and Record this meeting appears as it draws near. Your administrator can switch these assists off: the Tools then stay entirely manual.
Which model answers, and who can switch these assists off?
What you write and what a sentence asks go through the Tool Assist line of the LLM configuration, a message’s summary and sorting through Email Digest, writing a message through Emails Agent, an image or a scan through Vision Analysis: your administrator sets them. An answer already computed comes back at no cost.

An administrator switches the assists off in two places:
• Settings › Administration › Platform capabilities › LIA in the Tools, at once, with no redeployment
• the server configuration (TOOL_ASSISTS_ENABLED=false)

Once off, the lines served by LIA leave, Keep in a space and reading aloud included: what you do by hand, your reminders and your tickets stay.
Why can’t I do something my provider’s own app does?
Each screen offers only what your provider’s connection lets LIA do: a gesture one provider lacks is not offered rather than offered and refused. A page cut at a limit always says so.
How do the text menus work throughout LIA?
Select readable text anywhere in LIA, including chat, for Read, Question, Explain, Summarize, Simplify, Translate, Internet, Task, Event, Reminder, Ticket, Remember or LIA. Answers and audio stay in a compact bubble. Creation actions open the corresponding prefilled form for you to adjust and save; a reminder without a clear time needs completing. LIA opens chat with an unsent draft. While editing, select text for Translate, Correct, Format, Shorten, More formal or Warmer. Translation lets you choose a suggested language or enter another. Replacement always needs your click. Empty fields and collapsed selections have no contextual menu; dictation uses its separate controls. Moving focus to another field closes the editing menu. Internet names its actual search service and source links; unavailable capabilities cannot run. AI and voice use the usual usage limits and show their cost; a connected search service may bill your own key.

Phone Calls

LIA can place real phone calls on your behalf. Complete setup guide (Twilio + ElevenLabs) and everyday use.

What can LIA do on the phone?
LIA can place a real outbound phone call for you, pursue the goal you set, and report back in the chat.

Examples:
• "Call the bakery and ask if they have a cake for Saturday"
• "Call the garage to check if the car is ready"
• "Call Marie and ask if she's free Tuesday evening"

During the call, LIA introduces itself as an assistant calling on your behalf, holds a natural conversation toward your objective, then hangs up and writes you a summary — with a one-tap follow-up action when relevant (for example, creating the event that was just agreed).

💡 Good to know: LIA only places outgoing calls — it never answers your incoming calls.
How does a call work, step by step?
1. You ask in the chat
• Call Dr. Martin's office and ask for an appointment this week
2. LIA asks for your confirmation
  • Before dialing anything, LIA shows you a confirmation card: who it will call, which number, and why. Nothing happens without your explicit approval.
3. The call takes place
  • LIA's voice agent makes the call from your number and works toward the objective. You can keep chatting with LIA — or close the app — in the meantime.
4. The summary arrives in the chat
  • A written summary is delivered asynchronously, with a suggested follow-up action when there is something to do next. Your recent calls are also listed under Settings › Preferences › Recent calls, with their status, outcome, summary and cost.
  • The summary comes back to the discussion you asked for the call in. If the call reaches a voicemail, LIA tells you whether its message was left; if nobody answers or the line fails, it tells you that too, and invents nothing. If the provider never reports on the call, or does not answer the request to call, LIA says it does not know how the call ended rather than staying silent; a refusal that will not pass on its own is never presented as a passing problem.
What do I need before starting?
Phone calls use your own accounts with two external services — LIA adds no charge of its own:

1. 📞 A Twilio account — provides the phone number LIA calls from (number rental + per-minute pricing).

2. 🎙️ An ElevenLabs account — provides the voice agent that speaks during the call (requires access to the Agents platform; call minutes consume your plan's credits).

3. ✅ The feature enabled by your administrator — if you don't see a Telephony block in Settings > Connectors, ask your administrator to enable it.

The full setup takes about 15 minutes (plus possible validation delays for a French number) and follows 4 steps, each detailed in the next questions:
1. Get a phone number (Twilio)
2. Prepare ElevenLabs (API key + number import)
3. Configure the return webhook
4. Activate the connector in LIA
Step 1 — Get a phone number (Twilio)
Twilio is the operator that provides the number LIA calls from.
1. Create the account
  • Sign up at twilio.com and upgrade to a paid account (the free trial only calls verified numbers).
2. 🇫🇷 For a French number: submit a Regulatory Bundle
  • French regulation requires an identity document and a proof of address: Phone Numbers > Regulatory Compliance > Bundles. Human validation takes up to ~3 business days.
  • Easiest: a local number (01–05) or national number (09)
  • Avoid mobile numbers (06/07): extra requirements and 1–2 weeks of provisioning
3. Buy the number
  • Phone Numbers > Buy a Number — make sure the Voice capability is checked.
4. Note your credentials
  • On the Twilio Console dashboard, copy the Account SID and the Auth Token — you will need them in the next step.
  • 💡 Just testing? A US number can be bought instantly, with no regulatory bundle.
Step 2 — Prepare ElevenLabs (API key + number)
ElevenLabs provides the voice agent that speaks during the call.
1. Check your plan
  • You need a plan with access to the Agents platform. Call minutes consume your ElevenLabs credits (on top of Twilio's per-minute cost).
2. Create an API key
  • Profile (bottom left) > API Keys > create a key. If you restrict its permissions, it needs at least read/write access to the Agents platform (LIA creates its agent, lists your numbers, and triggers outbound calls). A full-access key works fine to start — you can restrict it after your first successful call.
3. Import your Twilio number
  • Agents Platform > Phone Numbers > Import from Twilio: enter the number in international format (e.g. +33612345678), the Account SID, the Auth Token, and a label.
⚠️ Two important points:
  • Use the native Twilio import, not a SIP trunk — LIA triggers calls through ElevenLabs' Twilio integration.
  • Do not create an agent yourself: LIA automatically provisions its own secured agent when you activate the connector (it will appear in your workspace as "LIA telephony — your name").
Step 3 — Configure the return webhook
The webhook is the channel through which the call report comes back into your chat. Without it, calls go out but no summary ever returns.
1. Open the webhooks settings
  • In the ElevenLabs Agents workspace settings, open the Webhooks section (only a workspace admin can do this).
2. Create a post-call webhook
  • Event type: post_call_transcription (transcription only)
  • URL: the address LIA shows you at step 3 of its wizard, with a copy button — it looks like https://your-lia-domain/api/v1/telephony/webhook
  • Authentication: HMAC
3. Copy the signing secret
  • ElevenLabs generates a signing secret when the webhook is created — copy it immediately, you will paste it into LIA in the next step.
  • 🔒 Leave "Send audio data" disabled: LIA never receives nor stores any audio — that is a design guarantee.
Step 4 — Activate the connector in LIA
Everything now comes together in LIA, in a 3-step wizard.

Go to Settings > Connectors > Telephony and click Connect. A notice reminds you that calls are billed on your own ElevenLabs/Twilio accounts.
Step 1 — API key
  • Paste your ElevenLabs API key and click Validate. LIA checks the key and lists the phone numbers of your workspace. ("No numbers" means the Twilio import from Step 2 is missing.)
Step 2 — Number
  • Pick the number LIA will call from.
Step 3 — Webhook
  • Copy the displayed webhook URL (if you haven't configured it in ElevenLabs yet, do it now — see the previous question), paste the signing secret, and click Activate.
  • LIA then creates its secured agent in your ElevenLabs workspace and the connector goes active. You're ready to call! 🎉
How do I place my first call?
💡 Recommended first test: have LIA call your own mobile — you'll experience a call from the other side.

1. In the chat:
• Call +336XXXXXXXX and ask whether the package has arrived

2. LIA shows the confirmation card (callee, number, objective) — confirm.

3. Your phone rings: the agent introduces itself as an assistant calling on your behalf, asks its question, and hangs up.

4. The summary appears in the chat a few moments later — asynchronously, so you can chat in the meantime.

You can also target a contact by name (“call Marie…”): LIA looks up the number in your contacts and asks you to disambiguate if several match. Your recent calls are listed under Settings › Preferences › Recent calls.
Is it private? Can LIA reveal my calendar or record the call?
Privacy is enforced by design, not by instructions:

• 🔒 Read-only, free/busy only — during a call LIA can check whether you're free or busy at a given time, but never reveals your event titles, participants, locations or any content. The call agent simply has no access to that data.
• 🚫 No recording — the call is not recorded.
• 🗑️ No stored transcript — the full conversation is never saved; only a short summary is kept, and it expires automatically (30 days by default).
• 📢 Transparent disclosure — the agent always introduces itself as an assistant calling on your behalf; it never pretends to be you.
• ✅ You stay in control — every call is confirmed by you before it is placed.
Can the assistant accept an extra or commit to an expense for me?
No, never. The assistant operates under a strict mandate: it can only act within the exact scope of your request. If the person offers an extra, an option, a price change or any unplanned commitment (even a small one, like a 3€ cheese topping on a pizza), it neither accepts nor declines on your behalf: it notes the exact offer and its price, explains that it cannot confirm by itself, and announces a call-back to confirm. You then find everything in the call summary — every cost with its amount, every open point — and LIA asks you how to proceed. If you wish, just request a new call with your instructions ("call back and accept the cheese topping").
How much does it cost?
LIA adds no charge of its own — costs go to your two external accounts:
📞 Twilio
  • Number rental: ~1–15 €/month depending on the type
  • Outbound minutes: per-minute rate based on the destination
🎙️ ElevenLabs
  • Agent call minutes, taken from your plan's credits
🤖 LIA
  • No surcharge, no markup — only the small LLM synthesis of your summary is counted: it is shown on the summary in the chat and in your LIA usage statistics.
🛡️ Built-in guardrails against surprises:
  • One active call at a time
  • Hourly call cap per user (10/hour by default)
  • Hard cap on call duration (10 minutes by default)
It doesn't work — what should I check?
"Invalid key" at step 1
→ The key is mistyped or lacks permissions (Agents platform access required).
"No numbers" after validating the key
  • → Your number isn't imported into the ElevenLabs Agents workspace, or was imported as a SIP trunk instead of the native Twilio integration.
The call goes out but no summary comes back
  • → The webhook is the issue: URL not publicly reachable, wrong signing secret, or wrong event type (it must be post_call_transcription).
"A call is already in progress"
  • → Only one active call at a time, by design. A stuck call is automatically cleaned up after ~15 minutes.
Calls refused after several in a row
  • → You've hit the hourly cap (10 calls/hour by default).
  • Still stuck? Ask your administrator: the server logs state precisely which webhook events were received and why any were ignored.
Can I follow a call while it is happening?
Yes. As soon as LIA starts dialling, a banner appears above the conversation: who she is calling, and where the call stands. It disappears the moment the call ends.

📒 Afterwards:
The history lives in Settings → Preferences → Recent calls: name, objective, status, outcome and recap. Never the number — the API does not expose it. The section does not appear at all if you have never placed a call.

⚠️ If the call is refused:
The message says so plainly, and distinguishes two cases: a transient failure (retrying makes sense) and a refusal from the carrier — unverified source number, exhausted credit — where retrying changes nothing until it is fixed on their side.

🔎 The facts, too: the debrief also lists the concrete findings the call produced — opening hours, availability, the answer you were waiting for — even when nothing needs doing afterwards.
Can LIA accept a date or an extra charge proposed during the call?
No, never automatically. The call summary first shows what the other party actually proposed — date, place, extra cost, pending decision — then, separately, the possible follow-ups: create a reminder, make a commitment, plan a meeting. Each one prepares the message in the composer instead of sending it, and every action that writes externally keeps its own confirmation prompt. No path turns “the other party proposed X” into “you accepted X”.
Can LIA call me?
Yes. Under Settings → My phone, type your own number (international format preferred) and ask for a verification call: LIA calls you, reads a four-digit code aloud, and you type it in. From then on, “call me”, “call me at eight to go over my day” or a routine that says it places a real call to you — without a confirmation card, because the person who would confirm is the one who picks up, and the number is proven yours.

🔒 The identity is the number, never a name:
LIA never calls a contact named “me”, and the third-party call tool refuses your verified number. If someone else answers, the assistant apologises, shares nothing and hangs up. Changing the number cancels the verification.
What does LIA know during a call with me?
It depends on the mode you chose in Settings → My phone (see the question below).

🎧 In Live (the default):
the voice knows nothing and reads nothing by itself: each thing you ask is handed to LIA, which runs it as a chat turn while you speak, with everything the chat knows — memories, agenda, reminders, open threads, your last exchanges — and answers through the voice.

🎯 In Live direct:
with Rich context on the phone on, the call carries what the chat would know; off, LIA knows only your availability. And if your administrator enabled live lookups, the voice reads your data itself during the call — the latest e-mails, a day of agenda, a contact, a file, a place nearby, the weather, what it retains about a topic — and reads the result to you in words. It reads and never acts: no message sent, no event created, nothing deleted. In My phone, each domain has its switch — the same switches govern a direct live session in the browser.
What happens after the call?
🎧 In Live:
nothing is left to relay — each thing you asked already ran as a chat turn of yours while you spoke, and the thread shows them, with a phone badge. Something that needed your confirmation is a draft waiting in the chat.

🎯 In Live direct:
what you said comes back as your own message: LIA writes it up, replays it in the chat as your turn — with a phone badge — and answers as it would have if you had typed it. Something you asked it to do becomes a draft waiting for your confirmation in the chat; the assistant never promises to have done anything during the call.

📣 And if it could not:
whatever the mode, the notification says exactly why — nobody answered, the line failed, someone else picked up, a question was already waiting for you, the conversation was busy, a ceiling was reached — and the call list shows the same verdict, next to the mode the call ran in.
How much does a call with me cost?
Two bills, one of them yours alone.

💶 What LIA counts:
the live lookups made during the call, the synthesis after it and the relayed turn are counted under one bill, shown on the relayed answer in the chat — or, when nothing could be relayed, on the call's report — and on the calls list, like any other request.

📞 What ElevenLabs bills:
the voice agent's own model, the speech, the line — everything that runs on your ElevenLabs key — is billed by ElevenLabs on your account and is never counted nor shown in LIA. Same rule for every service you use with your own keys.
Live or Live direct: how does LIA conduct my calls?
A choice of yours, in Settings → My phone, and the same two modes as the browser's live sessions.
🎧 Live (the default):
  • what you ask on the phone is handled by LIA while you speak, as a chat turn of yours — the voice announces it, keeps talking, and reads the answer when it lands. A confirmation LIA needs is asked on the line and settled by your next words; a question it asks is the answer. Everything is in the thread with a phone badge, counted under your limits.
🎯 Live direct:
  • the voice reads your data itself (the domains you allow) and acts on nothing; what you said is relayed to the conversation at the end as a message from you.
🛰️ When Live is not available:
  • Live needs the phone provider to reach your instance back; where it cannot (a private address), the setting says so and every call runs direct. Should the provider refuse the delegation on a given call, that call runs direct and the notification says so.
Does LIA's personality apply on the phone?
On a call with you, yes: after the opening sentence that checks it is really you, the voice speaks with the personality chosen in LIA Style — its words, its manner, its humour — as in a Live session, facts unchanged. On a call to someone else (a restaurant, a garage, an office), never: LIA calls on your behalf and speaks with a cordial, plain and efficient manner, the same for every account.

When your ElevenLabs agent's voice performs expression tags, LIA writes on the agent, before each call, the tags that fit that call: the whole palette for a call with you, the cordial ones only for a third party.
Can LIA place its calls through Fish Audio instead of ElevenLabs?
Yes: the telephony connector lets you choose between ElevenLabs and Fish Audio. One service at a time: activating one switches the other off. Everything LIA does around a call stays the same — the confirmation before calling someone, the report in the chat, calls with you in Live or Live direct.
🔧 Setting it up
  • In Settings > Connectors > Telephony, choose Fish Audio, paste your Fish Audio API key, pick the number LIA calls from, and activate. There is no webhook step: LIA creates the voice agent in your Fish Audio account and writes on it, itself, the address each call reports its end to. For that, your LIA server must be reachable from the internet.
📌 What differs
  • Publishing — at Fish Audio an agent runs the last version you published. After changing its voice, language or model on the Fish Audio site, publish it. LIA never publishes your changes for you: when a call needs it to and changes of yours are waiting, the call is not placed and LIA tells you what to do.
  • Answering machines — Fish Audio does not detect them: when the agent recognises a voicemail, it hangs up and leaves no message.
  • Numbers — the numbers Fish Audio sells are American and Canadian; for another country you bring your own line into your Fish Audio account.
⚠️ Good to know
  • This path was checked against the real Fish Audio service for everything except an actual phone call: no real call has been placed through it yet. Have LIA call you once as a test before relying on it.

Voice Mode

Talk to LIA and let it answer out loud. Push-to-talk, wake word, transcription languages, what leaves your device — and the Live mode, a real-time conversation on your own key.

How do I activate voice mode?
A hands-free badge (microphone icon) always sits in the chat header. Greyed, hands-free mode is off: tap it to turn it on — the same switch as Settings > Voice Mode. Once on, tap the badge to start speaking — or, with a French interface, say "Dis LIA" (wake word, in beta); hold it (or hold Space) to turn hands-free mode off.
How does push-to-talk work?
When voice mode is off and the text field is empty, hold the send button to record your voice. Release to stop — LIA transcribes your speech and places the text in the input field. Works on desktop and mobile.
Does voice recognition support my language?
Yes, if it is one of the six interface languages (French, English, German, Spanish, Italian, Chinese). LIA transcribes your voice in your account's language, chosen in Settings: it does not guess the language you speak, so to dictate in another of these languages, change your account's. The wake word itself only exists in French for now ("Dis LIA", in beta: its recognition is still improving); in the other interface languages, tap the badge to speak.
Is my voice data sent to external services?
Wake-word detection runs in your browser; the detection model does not upload the audio it listens to. Local speech-to-text uses Sherpa-onnx Whisper on the LIA server. Optional remote STT sends your recording to the speech engine your administrator configured (ElevenLabs Scribe, OpenAI or Fish Audio — the setting names it); it is off by default and requires your choice in Settings → Voice mode.

Other voice paths have their own boundaries: spoken responses can use your configured TTS provider, Live and Live direct send the session audio and useful context to the selected voice provider, and an enabled Simli avatar receives LIA’s output audio for lip synchronisation. Those services’ processing and retention policies apply. Disabling remote STT does not disable these separate paths.
Can I use a more accurate (remote) speech-to-text?
Yes. In Settings -> Voice mode you can switch from Local (default, free, Sherpa-onnx Whisper running on the LIA server, audio never leaves) to Distant (the engine your administrator configured — ElevenLabs Scribe, OpenAI or Fish Audio — billed by audio duration; the option names it). The remote mode is more accurate on conversational speech, especially in noisy environments. A discreet 🎤 badge on each user message bubble shows the duration and EUR cost of the transcription. The cost flows into your usual Cost dashboard tile and your usage limits - no separate quota. Two safety levers protect against cost spikes: a hard cap of 5 minutes per clip (admin-tunable), and a kill switch per provider (ElevenLabs, Fish Audio) the administrator can flip to force everyone back to local.
Which voices does LIA use for speech, and what does TTS cost?
LIA's voice synthesis is configured by the administrator from a single catalogue with five providers: Edge TTS (Microsoft neural voices, free, default), OpenAI TTS (tts-1 / tts-1-hd, premium voices: alloy, echo, nova, ...), ElevenLabs (eleven_multilingual_v2 / eleven_turbo_v2_5 / eleven_flash_v2_5, including custom and library voices from the configured account), Google Gemini, and Fish Audio (s2.1-pro, s2-pro, drama-3, with the vendor's licensed voices, the instance's own cloned or designed ones and the public voices an administrator pinned for every voice slot). Each person may choose LIA's male and female voice among the engine's voices in Settings → Voice mode. For paid providers, a 🔊 badge on the assistant message shows the character count and the EUR cost; Edge stays free with no badge. Latency-side: a sentence-level streamer starts playing the first sentence in about 1 second, while the assistant is still composing the rest of the answer.
Does LIA read numbers, prices and addresses correctly?
Yes. A full stop ends a spoken sentence only when a space follows it, so a decimal number, a price, a version number or a web address stays inside a single spoken phrase: "it's 3.5 degrees" is read as one sentence, not as "it's three" followed by "point five degrees" in a separate audio clip.

The rule is the same whether the response is spoken in one go or streamed sentence by sentence, so what you hear does not depend on how fast the answer was generated.
What is live voice mode, and how do I start a session?
A conversation with LIA voice to voice, in real time: you speak, it answers as it listens, you can interrupt it. It runs on a live model you connect with your own key — Gemini Live, GPT-Live, an ElevenLabs agent or a Fish Audio agent — from Settings → Connectors (the « Live » family, additive: you may hold several). Once a connector is active, the voice icon in the chat header becomes a menu: Live session (brand) opens one. A band above the thread shows the captions, the time left and an indicative meter; the extension dialog asks before the cap, and after a silence the session goes on standby instead of ending — nothing is billed until it wakes.

🧠 Two intelligences, one seam:
the voice holds the conversation; every request for data or action is handed to LIA, which runs it as an ordinary chat turn — the same confirmations, the same registers, the same quotas — drawn in the thread while you speak. A question LIA needs to ask you is what the voice says next.

📱 On iPhone and iPad: an ElevenLabs agent goes through its provider's native audio link (WebRTC), which is steadier. If Google refuses the connection because your Gemini Live key is restricted by IP address, LIA tells you: use a key without an IP restriction for this connector.
« Live session » or « Direct live session » — what is the difference?
🎧 Live session — the voice delegates: each thing you ask becomes a chat turn of yours, run by LIA while you speak, with confirmations for anything that acts and every answer drawn in the thread. This is the mode that can DO things.

🎯 Direct live session — the voice reads your data itself, without going through the chat: your agenda, mails, contacts, tasks, places, the weather, what LIA remembers — the domains you allow in Settings → My phone, the same switches as the phone. It acts on nothing: a request is noted, and at the end everything you said is relayed to the conversation as a message from you, which LIA answers. It buys latency, not economy: the tools' declarations are re-billed on every turn by the provider, about nine times the text prompt of a delegated session in our measurement. Offered only where the model's wire carries a tool schema (GPT-Live does not).
What does a Live session cost, and who pays?
Two bills, one of which is yours alone.

🔑 What the provider charges — on your key:
the live model's audio, in and out, is billed by the provider on your own account. The band shows an indicative meter folded from the provider's own usage reports and the tariff your administrator declared, with the context size; a cost that would need an undeclared rate is shown as unavailable, never as a partial figure. On an ElevenLabs agent the platform prices nothing: the clock alone runs, and the provider's own bill is shown once at the end. Nothing of this is recorded anywhere — it is your key. Settings › Live voice mode lets you set an optional spend ceiling per session.

💶 What LIA counts:
what LIA herself spends inside a session — each request you delegated, and in a direct session the lookups and the relayed turn — is counted like any chat message, under your usage limits, and added up on the closing card.
What does LIA keep of a Live session?
🎧 In a Live session:
each request you delegated is a chat turn, kept like any other, and the exchanges that stayed voice-only — small talk, an aside — are archived as two visible lines at the end of each turn, so the conversation reads as it happened; memories and interests learn from them as they would from a typed message. The session closes on one card: how it ended, the duration, the requests, the voice exchanges, LIA's own cost.

🎯 In a direct session:
nothing is archived while you speak — the captions live in the band. At the end, what you said is written up and relayed to the conversation as a message from you; the card says « relaying », then is rewritten once LIA answered, with its cost re-read. In a direct session, microphone audio goes from your browser to the provider. Without an avatar, the audio response returns through that same connection. With the avatar, that response also passes through the API and then Simli, without being recorded by LIA.

🧾 In every case:
the registers record that a session took place and which capability LIA read for you — never a word of what was said.
What happens when a Live session goes on standby?
After the model's silence timeout, when the page stays hidden, or when you tap the standby button, the session falls asleep instead of ending: the connection to the provider closes, so nothing is billed during the standby, but the session stays open — its requests, its meter and its closing card wait for you.

⏰ Waking it: tap Wake up in the band or, with a French interface, say "Dis LIA" (in beta) while the page is visible. LIA opens a new connection with its own context of the moment — the time, its state, your settings: the memory is the application's, not the provider's. On an iPhone, keep the screen on so that it can hear you.

🛑 Ending it: a session only ends on your gesture; an unbroken standby past the length your instance sets (8 hours by default) closes it by itself. The session's cap, its duration and the meter count awake time only, and the closing card says how many times it slept. In a direct session, what you said is relayed into the conversation at each standby.
How do I cut LIA's voice while it reads an answer?
By voice, with a French interface and hands-free mode on: say "LIA, stop" to cut the reading — nothing is recorded or sent to the chat — or "Dis LIA" to cut it and speak to LIA at once. Both phrases are detected in your browser, in beta: their recognition is still improving and they do not exist yet in the other languages. A "LIA, stop" said while nothing plays does nothing. In a Live session the provider hears you directly: simply speak to interrupt it.
Can LIA speak through an avatar, and what does it cost?
Yes, as an experimental beta, when your instance enables it. Connect Simli in Settings › Connectors using your own account key, stored encrypted, then explicitly enable the speaking avatar and choose a face. Connecting the service alone does not switch it on.

The avatar uses the audio LIA already produces for voice comments, Live and Live direct. Its floating window moves by dragging or arrow keys and offers three sizes. Radio keeps its own player. You may need to tap Enable audio when your browser asks for a gesture.

💶 Personal cost: the active connection consumes your Simli plan even during silence between answers. Live standby closes it; waking starts a new connection. Disable the avatar to close it. This charge belongs to your Simli account, outside the platform's usage limits.

If the avatar is unavailable, the existing text and voice paths remain usable. Speech-driven facial realism and physical Android/iOS behaviour are still being qualified; this beta does not promise identical results on every device.

Privacy: enabling it sends LIA's spoken output to Simli to generate the face. It does not send your microphone audio to Simli; the voice provider's own processing still applies.

The “Stop avatar session” button remains available in the window and shows whether closure is confirmed or still pending. If it remains pending, wait and try again; unconfirmed closure is not proof that Simli has stopped billing. The avatar, companion and shortcuts stay within the visible area when the mobile keyboard opens, you rotate the screen or zoom; your preferred position returns afterwards without these adjustments overwriting it.
Does LIA's personality also shape its voice in a Live session?
Yes — what the voice says, and how it sounds. The personality chosen in LIA Style governs how the voice addresses you, its words and turns of phrase, its humour, how it announces a lookup and reports what it found, and how it says goodbye, on every provider, in Live as in Live direct. Facts never bend to it: a result, a date or a failure stays exactly what it is.

🎭 With an ElevenLabs agent:
when the agent's voice performs expression tags (an Eleven v3 or v4 voice, or a realtime model), LIA writes on the agent, at every session, the palette of tags the voice may use — sarcastic, weary, soft, cheerful and others — and switches off the vendor's default persona; your personality decides which tags fit, and the tags never appear in the conversation. The portal's canned greeting is no longer heard: the voice's first words come from LIA. The voice, the language and the turn-taking stay those set on the ElevenLabs portal.
Do LIA's spoken comments follow its personality too?
Yes. When voice mode is on and LIA uses a tool, it speaks a short comment on what the answer kept. That comment is written in the personality chosen in LIA Style, from the facts of the answer and in your language: the way it addresses you, its words, its humour, how it reports what was found. An answer read aloud is the written answer itself, already in your personality.

🎭 Its tone:
with a voice engine that performs expressions — ElevenLabs v3 and v4, Fish Audio S2, Gemini 3.8 — the personality also picks how a sentence sounds: sarcastic, weary, soft, hurried, cheerful and others. The other engines carry it in the words alone. How strongly it comes through also depends on the model your instance uses to write the comment.
Can I hold a Live session with a Fish Audio agent?
Yes. In Settings → Connectors, the « Live » family offers Fish Audio beside Gemini Live, GPT-Live and ElevenLabs: paste your Fish Audio key, choose one of the agents of your Fish Audio account, activate. Both sessions work — « Live session » and « Direct live session ». The agent's voice, language and model are the ones you set on the Fish Audio site; LIA brings only its instructions and its tools.

📌 One thing to know — publishing:
at Fish Audio, an agent runs the last version you published. Before a session, LIA adds its tools to the agent and publishes that addition — only when nothing of yours is waiting. If changes of your own are waiting to be published, or if the agent was never published, LIA publishes nothing for you: the session does not start, and LIA tells you to publish or discard your changes on the Fish Audio site, then start again.

💶 What it costs:
the session runs on your own Fish Audio key, and Fish Audio bills it on your account. At the end LIA shows how long the session lasted; Fish Audio gives no amount there, so LIA shows none.

Personal radio

Your day and the world’s news, written and voiced for you while you listen — how to start it, tune it and keep its cost in view.

How does the personal radio work?
Your radio is a station made for you alone, while you listen: LIA opens with your day, then brings in the news from the outlets you follow, translated into your language. A fixed running order decides what comes next; models write the words and the voices you chose speak them, over the station’s own music.

📻 Where: the radio button in the header (on a phone, in the menu under the LIA logo), the radio card on your Home page, or the Radio page. Nothing is produced on a schedule, and nothing is kept for replay.

🔎 Every sentence has its source: Transcript and sources shows who speaks, what is said and the outlet or the part of your day each line rests on. Advertising and sales pitches are left out.

⚙️ Tune it in Settings › Radio (Preferences tab, Voice & Media): programmes, your life on air, news sources, voices, listening and spending. If your instance does not offer the radio, its page says so.
How do I start, pause and stop the radio, and when does it stop by itself?
▶️ Start: press Start the radio in the header, on the Home card or on the Radio page. The Radio page also sets this session’s Automatic stop (as in your settings, after a number of minutes, or none) and Listening in company. The station’s music starts right away and the bar tells you when the first voice should arrive.

⏯️ While it plays: a bar under the header follows you on every page — the station, what is on air, the minutes left, the cost so far, Pause, Stop and Transcript and sources. Your keyboard’s or headset’s play/pause key works too. Starting the radio again, even on another device, replaces the session under way.

⏹️ It stops by itself when:
• the automatic stop arrives — it counts listening time, so a pause pushes it back, and the station signs off with a short farewell
• nobody is listening: the page was closed, or the pause lasted too long
• its programmes repeatedly cannot be produced
• a spending limit is reached
The bar then says why and offers Start again.
Which programmes does the radio broadcast, and how often?
In Settings › Radio › Programmes, you set how often each programme comes back — Never, Rarely, Normally or Often:
• Personal journal: your day in three editions — what’s planned in the morning, what you did and what’s left at midday, a look back and the days ahead in the evening
• Headlines, News bulletin, Brief and The figure: always stories you haven’t heard yet
• Analysis and In depth with an expert, the station’s Editorial, The debate and Talking it through between its commentators: one story, from an angle — they may come back to a story you heard, never to the one the previous programme just told

The Opening and the Sign-off belong to the station. When LIA sends you a notification while you listen, a News flash breaks in, then the programme resumes where it stopped — unless you listen in company or switched off Notifications in your life on air.
What can the radio say about my own day?
In Settings › Radio › Your life on air, each source has its own switch, and a source switched off is never read, not merely left unsaid: appointments, reminders, tasks, commitments, workboard tickets, birthdays, unread and sent e-mails, what LIA did for you, the weather, health, meetings, LIA’s notifications, people to get back to, knowledge spaces, kept answers and the conversation under way. Each switch says what it reads.

Your day airs in the Personal journal: set that programme to Never and nothing personal is read at all. With Listening in company — for one session on the Radio page, or by default under Listening — nothing about you airs: only the weather is read, for the opening.

What the radio opened appears in the Consultations tab of your Registers, like any other reading. Your radio settings and sites are exported and deleted with your account.
Where does the radio’s news come from, and can I add my own sites?
A shared newsroom reads public feeds for the whole instance, and only while someone uses the radio. In Settings › Radio › News sources:
• The radio’s sources: major outlets and independent voices in several languages, all ticked by default — everything airs translated into your language. Untick what you don’t want to hear. Each shows how many stories it published recently, how many you haven’t heard, and Not answering when its last readings failed.
• Your sites: type an address, press Check — the radio finds the site’s feed and describes it — then Add. A site can be paused (kept, no longer read), renamed (the name heard on air) or removed, up to a number your administrator sets. A site that refuses robots or publishes no feed is refused, with the reason.

Advertising, sponsored content, deals and purchase pitches are left out; genuine invoices, reservations and editorial, scientific or economic reporting stay eligible. When an outlet offers no readable article, the station uses only its title or summary and never pretends to have read more.
Can the radio bring me news about my interests?
Yes, if you have your own Brave Search or Perplexity key in your connectors. When a session starts, your strongest interests — as many as your administrator allows — are searched with your key (Brave first, Perplexity otherwise), and the stories found can go on air. The news programmes then take turns between your sources and what the search found.
• The search runs on your account with that provider, so LIA does not record that spend; each search appears in your Consultations register.
• An interest searched recently is not searched again at every start.
• With Listening in company, or when your administrator has switched interests off, none of your interests is read, searched or mentioned.

Your interests and remembered preferences also help the news programmes pick their stories — as context, never stated on air as facts about you.
How do I name my station and choose its personality and voices?
In Settings › Radio › Voices:
• Station name: said on air by the host and shown on the player; leave it empty for “LIA Radio” in your language. A session keeps the name it started with.
• Station personality: The same as in the chat, or another of LIA’s personalities.
• A voice for each role — LIA, the Anchor, the Expert, the Editorialist — and for the three Debate and discussion commentators, Automatic by default. A commentator never shares a voice with LIA, the anchor or another commentator: those voices are greyed out.

Choosing a voice plays a short sample in your language; a radio that is playing pauses so you can hear it (a meeting recording or a Live session keeps priority). The voices are those of the engine your administrator set for the radio, shown by default in your language and multilingual ones; your choices are kept when the administrator changes model at the same provider. On engines that voice several speakers, a debate is recorded in takes, never with more voices per take than the engine accepts. In a discussion, a debate, the editorial, the opening or the sign-off, a voice may open its line with a chuckle, a laugh or a sigh when it can perform them — twice per programme at most by default (a bound your administrator sets), never during the news.
What does checking by a second model do, and why is it off by default?
Every programme already passes an automatic editor, at no extra cost: each line must cite the facts it rests on, a line stating a figure no cited source gives is removed, and a programme that loses too much is not aired.

Checking by a second model (Settings › Radio › Listening) goes further: before a programme airs, another model compares its sourced lines with the facts they cite, removes the unsupported ones, and may drop the programme if too little remains or the check fails.
• No check · Check the news · Check every programme
• the settings say how many kinds of programme your choice has re-read

It is off by default because every reading is billed and can delay the listening; it stays one choice away. An administrator can set another default for accounts that never chose — the settings always show the level that applies to you.
Can I read the articles the radio talks about?
Yes. On the Radio page (or Transcript and sources on the bar), the programme on air is shown line by line — the line being spoken is highlighted — with the outlet and date each sentence rests on, and what comes next.

Below, Articles of this session lists every story cited, in the order it aired. Open one to read it whole, in your language:
• it is translated when you open it, billed once, then read again at no cost — its cost is shown
• Original opens it on the outlet’s site in a new tab, with no translation asked
• when the outlet shares only a summary, or the article is long, the page says so and points to the outlet
• past your radio’s spending limit, it is shown in its original language, untranslated

The list stays once the session is over, so you can finish what you opened, until you start the radio again.
What does the radio cost, and how is it limited?
Each programme is written and voiced for you, so it is billed to your account and counts in your usage like any exchange: the writing, the speech, the optional model checks, the translations of the articles you open and the voice samples. The shared newsroom costs you nothing, and searches made with your own Brave or Perplexity key are paid to that provider, not recorded by LIA.

💶 In view while you listen: the bar shows the cost so far and, once enough has been produced, an estimate for the listening you planned — or per hour, with no automatic stop.

🛑 Limits: your account’s usage limits apply, and your administrator may also give the radio its own budget over a rolling 24 hours, shown in Settings › Radio › Your radio’s spending. Past it, a start is refused, a session stops before its next programme and articles stay untranslated — and the radio tells you when it can play again.
Why does the radio sometimes play only music, or never repeat a story?
The station’s music plays throughout, from the start to the farewell, in a mood that suits the hour or the news, and steps well back under every voice. You hear it alone:
• while the first programme is prepared — the opening is made first, from the time and the weather only, so the first voice comes quickly
• for a few seconds between two programmes, and while the next one is being written
• when a programme had nothing to say, or the editor refused its script: it does not air and the station moves on (a debate the voice engine cannot cast with distinct voices is not planned either)

🔁 No repeats: a story you heard is not told again, even by another outlet or in other words — only the angle programmes may come back to it. When everything has been heard, the station says so once (Nothing new). To hear everything again, use Forget what I heard in Settings › Radio › News sources: it asks first and cannot be undone.

Image generation

Generate and edit images using AI

How do I generate an image?

Simply ask the assistant to create an image. For example: "Generate an image of an astronaut cat". The image appears as a card below the response.

Can I edit a generated image?

Yes! After generating an image, ask the assistant to modify it: "Make it more realistic". The assistant automatically uses the most recent image.

How do I configure quality and size?

Go to Settings > Preferences > Image generation. Choose a quality, a size (square, portrait or landscape — with its resolution, 1K or 2K when the model offers both) and a format (PNG, JPEG or WebP). The choices depend on the image model your administrator configured; a saved choice it does not offer is replaced by the closest one it does, never by a dearer quality.

How much does image generation cost?

It depends on the image model your administrator configured: each quality in Settings > Preferences > Image generation shows its price range per image. An edit can also bill the image it starts from, when the model prices it per image. The exact cost of each image appears in the debug panel.

How do I download a generated image?

Three options: (1) Click the download button (arrow icon) that appears when hovering over the image — on mobile, the button is always visible. (2) Click the image to open full-screen, then use the download button in the top-right corner. (3) On mobile, long-press the image to use the native "Save Image" menu.

Is there a limit to how many images I can generate?

Two safeguards apply. Your account's usage limits (cost caps per billing cycle, set by the administrator) cover image generation costs like any other cost. On top of that, a technical rate limit protects against runaway loops: by default, 10 image generations (and 10 edits) per 5 minutes per user. Normal use is unaffected — if the limit is reached, the assistant tells you how long to wait before trying again. The administrator can tune the threshold and the window via the IMAGE_GENERATION_RATE_LIMIT_CALLS and IMAGE_GENERATION_RATE_LIMIT_WINDOW environment variables.

How long does a generated image stay available?
Not indefinitely — unless you keep it. A generated image is an attachment with an expiry, and a scheduled cleanup removes expired attachments every few hours.
🕒 What the card shows:
  • The deadline returned by the server, and an alert tone in the final hours. A kept image says it is kept; an image that no longer exists says so instead of showing a date.
Why not a hard-coded duration:
  • The retention window is configurable by the administrator, so a value written into the interface would eventually be wrong. Images that predate this feature show nothing rather than guess.
💡 Tip:
  • To keep an image past its deadline, keep it from Settings › My generated files (the pin on its card): a kept image is never deleted automatically, within your account's limit shown there. You can also download it.
Where do I find every image LIA has created?
In Settings › My generated files › Images: all your images, across every conversation — those a connection shared with you included — with a search on the name, two date filters and four sort orders.

Every card shows when its image expires, and the tone warms in the final hours. To keep an image past that deadline, keep it from its card: a kept image is never deleted automatically, within the limit your account shows above the gallery. You can also download it.

Clearing a conversation does not take its images with it: a reset removes only the attachments you sent yourself.

🗣️ Or ask LIA: "show me the lighthouse image from this morning" — it finds the image in your gallery and shows it again in the chat.
Which image models can LIA use?

The one your administrator configured, from three families: OpenAI's GPT Image, Qwen Image 3.0 (Alibaba Cloud) and Google's Gemini Nano Banana 2.1. All three create an image and edit a photo you attach.

Each model declares what it can do — qualities, sizes, resolution — and LIA offers you only that: if your preference does not exist on the chosen model, it takes the nearest option, never a dearer quality. The format you chose (PNG, JPEG or WebP) is applied to every image, whatever the provider, and the displayed cost also counts what the provider bills beside the image: the source image of an edit when it bills it and, with Gemini, your description and the model's reasoning, billed as tokens.

What does the “Enhance prompts automatically” option do?

It lives in Settings > Preferences > Image generation and is off by default. When it is on, before each new image a dedicated model rewrites your description following the advice the image provider publishes — framing, light, lens, style — without changing what you ask for.

What is protected: a text you put in quotes (a title, a sign) must come back unchanged in the rewritten version; otherwise, or when the rewrite fails, exceeds the allowed length or comes back empty, your original description is sent. The option therefore never blocks an image.

What it leaves alone: editing a photo keeps your request as is, and your gallery shows your own words. The rewrite is one short extra call, counted in your usage like everything else; your administrator can also withdraw the option for the whole instance.

Is image generation enabled by default?
Yes for your account, as long as your instance offers it. The Enable image generation switch, in Settings › Preferences › Voice & Media › Image generation, is on from the start. Turn it off and LIA no longer creates or edits images for you: it tells you the option is off in your settings.

Your administrator decides above that: image generation can be switched off for the whole instance at any time, without a restart. The qualities, sizes and formats you can choose depend on the image model your administrator configured.

Security and Privacy

Protecting your data

Are my conversations private?
Your account has controlled access, and conversations are isolated between users.

Storage: connector credentials and designated fields, such as saved locations, are encrypted. This does not encrypt every stored field: conversation messages are stored as database text, and the instance operator may have technical access.

Providers: LIA sends the context needed for a request to the AI and service providers configured for your instance. Their retention and possible training policies apply. Review those policies, provider settings and your hosting arrangements before sharing sensitive information.
How is my data protected?
Several controls work together:
• Account authentication, OAuth with PKCE and server-side sessions with HTTP-only cookies
• CSRF protection, revocable sessions and audit records of sensitive actions
• HTTPS when configured by your operator, and encryption of credentials and designated stored fields
• Access controls between accounts and filtering of personal information in operational logs
• Context sent to configured providers only as needed for the selected features, under their own policies
🌐 Reading the web:
  • Every address a page sends LIA to is checked before connecting, redirects included, one hop at a time
  • The connection goes to the address that was checked: neither your internal network nor the server's can be reached this way
  • A booby-trapped text (e-mail, page, message) is formatted in time proportional to its length and cannot stall the server
💾 Backups:
  • The deployment provides daily backup and restore-check tools. Your operator must enable and monitor them; recovery depends on the latest successful backup and the chosen retention policy.
🧬 Software supply chain:
  • Every release ships a locked, cryptographically verified list of components (reproducible builds)
  • All dependencies — including indirect ones — are automatically scanned for known vulnerabilities
What data does LIA collect?
LIA stores the data needed for your account and the features you enable: profile and preferences, conversations, memories, selected documents, connected-service credentials and usage records.

Your home address and, if you opt in, your last known position are stored encrypted. Each position update replaces the previous one; disabling location memory deletes it.

Connected services are read under the permissions and features you authorise. Useful excerpts can appear in conversations or documents you keep. You can export your data and request account deletion; accounting retention and backup policies are explained below.
Can I delete my account and data?
You can ask your administrator to delete your account.

Personal content: deletion purges conversations, memories, connected-service credentials, documents and the other personal records covered by the account-deletion flow, including health data and saved locations. Purged content cannot be recovered through reactivation.

Accounting and retained records: billing records remain. Deletion can keep your name and email for billing contact; final erasure deletes the account row. Audit or diagnostic records may still contain identifying details under their retention policies. Backup copies follow the operator’s retention policy.

Disconnecting a service or clearing your conversation is a narrower alternative when you want to keep the account.
Does LIA have access to all my emails/files?
Access follows the permissions of each connector and the features you enable.

LIA can read useful emails, files and calendar events for a request. Authorised scheduled actions, briefing, proactive features and Radio may also read their selected sources without a new chat message.

Documents you choose to import or synchronise are indexed in your knowledge spaces. Service results may be cached, and excerpts used in an answer can remain in conversation history or a document you keep.

Review your connector permissions and enabled features; you can disconnect a service, pause a knowledge space or delete its imported data.
Do external services see my data?
Yes, when a feature needs them.

AI providers receive your question and useful context, such as selected messages, memories or document excerpts. Speech, image and Live services receive the text, audio or images required by their respective modes. The optional Simli avatar receives LIA's spoken output after your explicit opt-in.

Connected services receive the requests needed to use them: searches, requested locations or actions you confirm. Weather services can receive coordinates, not just a city name.

Each provider applies its own privacy, retention and possible training policies. LIA's access controls and local storage do not replace those policies; check the configured providers and share only the information needed.
What are passkeys and how do I enable them?
A passkey lets you sign in with Face ID, your fingerprint or your device code — no password typed, nothing to remember.

🔑 Why it is safer:
• Phishing-resistant: a passkey only works on the real LIA site
• LIA receives the public key, never the private key; your passkey manager controls any device synchronisation
• A leaked password becomes useless to an attacker

⚙️ How to enable:
1. Go to Settings > Security > Strong authentication
2. Click Add a passkey and follow your browser (Windows Hello, Face ID, Android…)
3. Optionally name it ("iPhone", "Work PC") and add one per device

On the login page, your browser will then offer the passkey directly in the email field (or use the "Sign in with a passkey" button).
How do I add a second verification step (TOTP)?
The authenticator app adds a 6-digit code as a second step when you sign in with your password or with Google. A passkey sign-in does not ask for it: a passkey is already a strong proof on its own.

⚙️ Setup:
1. Settings › Preferences › Strong authentication › Authenticator app (TOTP) → Enable (you may first be asked to confirm it's you)
2. Scan the QR code with Google Authenticator, 1Password, Aegis… — or enter the secret manually
3. Type the 6-digit code the app shows, then Activate

🔒 Backup codes: 10 single-use codes are shown exactly once — store them safely; each one replaces the app code if you lose your phone. Regenerate backup codes gives you a fresh set at any time (the old set stops working).

Signing in then takes two steps: your password or Google, then the current code from your app. The section appears when your instance offers strong authentication.
Why does LIA sometimes ask me to confirm my identity?
Sensitive actions are protected by an identity confirmation (a "step-up"):

🛡️ Protected actions:
• Managing passkeys and the authenticator app
• Exporting your data
• Signing out all other devices
• Disabling password sign-in

⏱️ How it behaves:
• Right after signing in, nothing is asked for 5 minutes
• Afterwards, one quick confirmation (password, code, passkey — or a fresh Google sign-in for Google accounts) re-opens the window

The point: even if someone borrows your open session, they cannot quietly lock you out or walk away with your data.
How do I see and disconnect the devices connected to my account?
Settings > Security > My devices lists every live session of your account.

💻 What you see:
• Browser and system families (e.g. "chrome · windows") — never the full technical detail
• A truncated IP (e.g. 192.168.1.x) and coarse last-activity time
• The device name when it is registered for push notifications

🔌 What you can do:
• Sign out one device — it loses access immediately, even mid-conversation
• Sign out all other devices (asks for an identity confirmation)

🔔 New sign-in alerts: your devices receive a push notification when the account signs in from an unrecognized device — the toggle to turn it off is in the same section.
How do I export all my data (GDPR)?
You can download a complete archive of your data at any time (GDPR right to portability).

📦 What the archive contains:
• Your profile and settings
• Conversations, memories, journal — as readable Markdown AND structured JSON
• Your uploaded files (attachments, knowledge-space documents)

🚫 What it never contains: connector credentials, security keys, push tokens — secret material is unexportable by design.

⚙️ How: Settings > Security > Data export → Request an export → a push notification tells you when the ZIP is ready (downloadable for 24 h).
What is left on the device when I sign out?
Nothing that belongs to you. Signing out does not just close the session: it clears what the browser and the device were holding on your behalf.
🔔 Notifications:
  • This device's push registration is revoked before the session closes
  • On a shared computer or phone, the next person will not see any of your notifications
📍 Location:
  • The last known position is erased
  • So is the geolocation permission — the next account has to give its own consent, it does not inherit yours
💬 Content:
  • The message draft in progress is deleted
  • No access token was ever stored in the browser (BFF architecture), so there is none to remove
  • The same cleanup runs if another account signs in through the same tab without a sign-out first — after a session expires, for instance.
Can an email or a web page give LIA orders?
No. A text LIA reads is never a text LIA runs.
📨 The problem:
  • Every day LIA reads content you did not write: an email body, an invitation description authored by its organiser, a web page, a place listing, an external server's result. Anyone can slip in a sentence like "ignore previous instructions".
🏷️ The protection:
  • Every piece of data carries its provenance. The 25 data types LIA handles are classified once and for all: produced by LIA itself, or written by a third party. An unknown type is treated as external for safety, and the application refuses to start if a type has not been classified. What comes from outside reaches the model tagged as material to analyse, never as an order to follow.
🔎 The detection:
  • Seven families of trap are recognised across the app's six languages: fake system message, instruction hijack, identity switch, requests to send your data elsewhere, the name of a LIA tool slipped into foreign text, invisible characters, a directive hidden in an HTML comment.
🧠 And the memory keeps the label:
  • When a long conversation is summarised to save room, provenance used to be lost with the detail — a sender's request could resurface later as if it were an established fact. The summary now carries the label with it: what a third party wrote is reported in its own section, attributed to its source, and is never restated as one of your decisions or as something to do.
✋ What LIA does not do:
  • It never rewrites your content. An email stays exactly what its author wrote; only a note is added beside it. Cleaning the text would give the illusion of a guarantee that the next bypass would deny — and would alter a message you may have wanted to read as-is.
Can I get the complete log of what LIA did for me?
Yes, and complete means complete. Open the Transparency registers page — its Registers shortcut is on the Home page:
• the Actions and Consultations tabs each take their whole register out as Export (readable) (Markdown), Export (CSV) or Technical (pseudonymised JSON Lines)
• below the tabs, Unified extraction gathers all five records — your turns, the actions, the consultations, the model calls and any gap in the record itself — into one JSON Lines file with no content and every identifier replaced by a handle, so you can attach it to a request or a complaint as it is

Nothing is truncated, whatever the volume: each file states the exact count (per record in the unified file) along with the moment the extraction was taken. The download is compressed on the fly when your browser accepts it, without the file name changing.

🗣️ Or just ask:
• What did you do for me this week?
LIA reads the same registers and answers with exact totals: what it did, what it consulted, and what it did on its own initiative.
Does a script written by LIA ever see my API keys or send my data somewhere?
No key, and no data without your word.
🔑 Keys
  • the script only ever sees a temporary token, valid for that single run
  • a dedicated proxy swaps it for your real key, on the host of that connector alone — the key never enters the container the script runs in
  • LIA's own account keys and your sign-in credentials are never offered to a script
📦 Data
  • a script that declares no host runs fully offline, with the data of the turn on its standard input
  • a script that wants a host you never allowed is stopped, and you choose: with the data, without it, or not at all
  • “Without” means the script receives nothing of what LIA gathered — it can only reach the host
📖 Traces
  • every network run is claimed in your register before it starts, with the hosts it declared
  • the script's code never appears in the answer: it is shown in the chat's debug panel, on a wide screen — to an administrator who turned the panel on, and to you if the administrator lets users open it and you switch on Enable debug panel in Settings › Preferences › Debug Panel
In a Live session, what does the provider see and what does LIA keep?
🔑 The provider, on your key:
your voice and LIA's answers transit directly between your browser and the live provider you connected with your own key — without the avatar, the audio uses this direct Live path. With the avatar enabled, LIA’s spoken output also crosses the API-controlled Simli relay. The provider receives the mandate LIA renders for the session (how to behave, what to hand over) and, in a direct session, the tools' declarations and the results it reads for you. Its own usage and its bill are yours and shown to you once; the platform records none of it.

🧾 LIA, on your account:
each request you delegated is an ordinary chat turn — kept, confirmed, counted, filed in the registers like a typed one. The voice-only exchanges of a Live session are archived as visible lines; in a direct session nothing is archived while you speak, and your words are relayed at the end as a message from you. The registers record the session and the capabilities read, never the audio nor a transcript of it. Deleting your account removes all of it.

🎭 Optional speaking avatar: after your explicit opt-in, LIA's spoken output is also sent to Simli, on your personal connector, to generate the face. Your microphone audio is not sent to Simli. Disable the avatar to close that connection; its active time, silence included, uses your personal Simli plan.
Can signing in with Google activate or unblock an account?

No. Signing in with Google proves an email address, nothing more: an account's status never changes on that occasion. A registration awaiting approval stays pending, and an account an administrator blocked stays blocked.

LIA believes the announced address only when Google vouches for it. If someone had created an account with your address without ever verifying it, your Google sign-in proves the address is yours: that person's password and sessions are revoked, and the administrators are notified as after an email verification. A deleted account is never revived.

And even an inactive account can always sign out.

What do the server's technical logs keep?

Facts, not your words. To run and to be repaired, a server writes technical logs; LIA's keep counts, lengths, identifiers and error codes — not your searches, your interests, the names of your contacts or files, nor the pages you visited.

Measurements too: the metrics and traces used to monitor the server name the page requested by its pattern — “a relationship’s card” — never by its content: a name or a search sitting in an address does not get into them.

Even when something fails: a database that refuses a value usually quotes it in its error message; LIA removes that quotation before writing the line and describes the error by its facts — which constraint, which table. Service access keys are masked in every logged address.

Checked continuously: an automated test reads every log line in the code and refuses one that would write a person's text. Two limits remain, stated as they are: an external service's error message in an unknown format can still quote a value, and the debug level — off in production — keeps more detail.

How does LIA stay protected against flaws in its dependencies?

Through a watch and rules that cannot be forgotten. Every week, an automated check reads the security advisories each of LIA’s dependencies publishes — including those no public database relays —, the announced end-of-life dates and the state of the browser engine. Every finding is fixed, or accepted in writing with a reason and a review date: never ignored in silence.

Careful updates: a new version comes in only after a cooling-off period — from a few days for a fix to two months for a major version — long enough for a possible defect to become known. An update never moves a dependency backwards, and a security fix becomes a floor nothing can go below.

A verifiable build: every building block of the server — images, languages, tools — is pinned by its digest, every version publishes the exact inventory of what it contains (a CycloneDX SBOM), and the code goes through the secret detector before every push. Before it is published, every version is installed on blank machines.

Do private inputs needed by a command pass through the model?

Keys, file choices, account connections and browser consent use the operation’s secure interface, rather than parameters invented by the model. A command is tied to the account and its authorised context; permissions and the resource version are checked again before writing. Encrypted command content does not hide the whole service from its operator: messages and operational data retain the limits described here.

I forgot my password — and what if I sign in with Google?

On the sign-in page, “Forgot password?” sends a link to your address. The link works once, for a limited time, and stops working as soon as your password changes.

If your account signs in with Google and has no password, there is nothing to reset: you receive an email that says so — sign in with Google as usual. If you turned password sign-in off for your passkeys and do not sign in with Google, the link stays your way back should you lose them. An administrator account is never reset by email.

For your safety, a new email only leaves after a short delay, and a few times a day at most: if you asked very recently, use the last email you received. These limits protect your account whoever makes the request, so nobody can flood your mailbox with reset emails.

Why might an anti-robot check appear when I sign in or sign up?

An instance can protect its sign-up, sign-in and password-reset pages with Cloudflare Turnstile, so that scripts can neither create accounts nor flood mailboxes. Most of the time you see nothing: the check runs by itself and appears only when Cloudflare asks for an interaction.

It reads technical signals from your browser in a frame served by Cloudflare; LIA receives only a single-use token, which its server checks, and keeps nothing of those signals.

If Cloudflare does not answer the server, you can still sign in, while creating an account or asking for a reset email waits until it answers again. If the check does not load in your browser — an extension blocking challenges.cloudflare.com, for instance — allow that address on this site.

Transparency registers

What LIA did for you, what it looked at, and how to check and export it.

What are the transparency registers, and where do I find them?
They record, automatically, what LIA does for you. Open them from the Registers shortcut on the Home page (under My shortcuts), or from the Action journal button of the Alerts and Activity pages. The Transparency registers page has four tabs:
• Actions — what LIA did on your behalf: an e-mail sent, an event created, a file deleted, a document generated, a notification sent
• Consultations — what LIA looked at to answer you: your e-mails, your calendar, the weather…
• LIA's own initiative — what it did and read without being asked
• Overview — the same records as charts

Actions and consultations are kept apart on purpose: they count different things and their totals never add up — one exchange can consult five sources and change nothing. Behind them, LIA also records each conversation turn, each model call and any gap in the record itself; you see those in the charts and in the complete extraction.
What does the “LIA's own initiative” tab show?
What LIA did without being asked: its proactive passes, the interests it explored for you, the notifications it sent unprompted, the discussions it started. The tab holds two lists, Actions taken on its own and Sources consulted on its own.

Everything you set in motion stays in Actions and Consultations: what you typed or said, what you asked in one click from Tools, your reminders, the routines you configured — a routine is your own instruction, deferred — and the sub-agents you delegated to. Opening your Home page briefing counts as your request too.

Why a separate tab? A background pass runs on its own schedule and can far outnumber what you asked for: mixed together, the few lines that matter would be lost. Every line also states its origin: You asked, Scheduled, Sub-agent or LIA's initiative.
Can LIA claim it did something it did not do?
No — by construction, not by promise. An action is written into the register before it happens, and marked Done only when the service that performed it returned an explicit success. The absence of an error never counts as a success.

🔎 What you can rely on:
• a failed action is recorded as Failed; one that could not be attempted because a confirmation was missing, as Refused
• an action you confirmed and that was already performed is never performed twice: LIA hands back the result recorded the first time
• if an action you must confirm cannot be recorded, it is not performed; any other action goes ahead and the missing line is counted in Gaps in the record
• a notification is an action too: recorded before it leaves, marked done only when its delivery reported success

The other states — In progress, Interrupted, Interaction requested — say exactly where an action stopped.
What exactly is kept about a consultation or an action?
👁️ A consultation: the capability and its domain (Emails, Calendar, Weather…), when, how long it took, whether it Answered or gave No answer, and who set it in motion. Never what was searched for: “consulted your e-mails” is a record, “searched your e-mails for Marie” would be a second copy of your data. Identical consultations in a row are shown as one line with a count (×3).

Not counted as consultations: an answer served from cache — your mailbox was never opened — and a Home page section you hid. A source that refused shows No answer, never a silent success.

✋ An action: a short description in your language (for example: Sent an email to Marie), its outcome, when, who set it in motion and, when you had to confirm, which kind of confirmation you gave — shown in the exports. The description and the service's reply, kept so a confirmed action is never performed twice, are stored encrypted. Neither register copies the words of your request.
What do the charts of the Overview tab show?
Ten charts over your whole history, computed on the server from the records themselves — nothing is counted in your browser:
• Activity per day, in conversation turns
• How turns ended and Execution mode (Pipeline or ReAct)
• Actions by outcome and Consultations by domain
• Average latency per tool
• Calls per model — the model that actually answered, not the one configured today — Calls per graph step and Tokens per model
• Gaps in the record, a chart meant to stay empty

Each chart shows the exact total beside its bars, including what was folded into “Other”, so you can check that they add up. A chart of averages is labelled “Average”, never “Total”. Sub-agents and external MCP servers are grouped under one word, because their names are not LIA's to publish.
How do I export my registers, and what does each file contain?
📄 One register at a time: the Actions and Consultations tabs each offer three downloads of the whole register, LIA's initiative included:
• Export (readable) — a document to read, with your own wordings
• Export (CSV) — a spreadsheet to count
• Technical — JSON Lines: the same events with no content at all and pseudonymised identifiers

🗂️ Everything in one file: below the tabs, Unified extraction gathers the five records LIA keeps about you — turns, actions, consultations, model calls and gaps in the record — one line per record. It follows the same contract as an administrator's extraction, limited to your account by construction: no content, and every identifier, yours included, replaced by a handle. You can attach it to a portability request, a complaint or a bug report without editing it first.

✅ Complete: no export is cut short; its header states the exact count and the moment it was taken.
How long are the registers kept, and can I delete them?
They are kept as long as your account — there is no automatic purge — and they leave with your account archive (Settings › Preferences › Security › Data export), together with a short sealing statement when sealing is on.

🗑️ No “clear” button: a record you could erase at will would prove nothing. Deleting or resetting a conversation removes its words, not the register lines: the turn stays as a dated line that no longer points at any text.

🔚 Deleting your account removes the registers, their seal and the record of gaps. One technical record remains, kept for billing disputes: the line of each model call — model, provider, settings sent, tokens and cost — with no content and no request.
Can I check that my registers were not altered?
Yes, when your administrator has switched sealing on — it is off unless they do. Each account then has its own chain of cryptographic fingerprints, and a card above the tabs says how far your registers are sealed and how many lines are not sealed yet. Verify integrity recomputes the sealed lines and answers Intact, or names where a break was found — report a break to your administrator. Note the final fingerprint it shows: comparing it later reveals even a rewrite made directly in the database.

✅ What it proves: a sealed line has not changed since it was sealed.
⚠️ What it does not:
• that a line was true when written — the recording rules answer for that, not cryptography
• anything about lines not sealed yet: sealing runs shortly after the fact, and the card counts them
• what older lines looked like before sealing was switched on: they are vouched for only as they stood at that moment
• a complete rewrite of the chain, unless you compare with a fingerprint you noted earlier

When sealing is off, the card says so: your registers stay complete and readable; only the proof is missing.
Why does LIA keep these registers?
Because an assistant that acts on your accounts should be able to tell you what it did, and what it looked at to do it — from a record, not from memory.

The recording is automatic: every capability that can act or read passes through the register by construction, whatever the route — the chat, Telegram, a routine, a sub-agent, a background pass, a voice session — so nothing depends on someone remembering to log it.

Article 12 of the European AI Act sets record-keeping requirements for high-risk AI systems; it does not by itself make this personal assistant high-risk. LIA's registers are a deliberate transparency measure — not a certification, nor a claim that every obligation of that article applies to every installation.
Can I ask LIA what it did, and how reliable is the answer?
Yes. LIA reads the same Actions and Consultations registers and answers with exact totals, even when it lists only the most recent lines:
• What did you do for me this week?
• What did you do on your own since Monday?
• Did you send the email to Marie yesterday?
• Which of your actions failed this month?

It can narrow by period, by who set the action in motion — you, or LIA on its own, exactly as the tabs separate them — and by outcome. Without a period, it covers the last few days. Each action is worded from the description recorded at the time, in your language, so the answer cannot drift from the screen.

It does not read the turns, the model calls or the Activity page: for those, open the registers or the activity timeline.

Self-hosting

Running LIA on your own server: the guided installer, what it protects, and what stays in your hands.

Can I run LIA on my own server?
Yes. LIA is free software (AGPL-3.0) built to be self-hosted, and a guided installer sits at the root of the repository: clone it, then run ./install.sh.

🖥️ What the machine needs:
• Linux, on a 64-bit x86 or ARM processor — a Raspberry Pi 5 is the reference production machine
• Docker with a recent Compose v2, and a recent Python 3 to run the installer itself
• at least 10 GiB of free disk space
The installer checks each of them and says what is missing.

Your conversations, memory, documents and encrypted keys live in your own database; the AI models and the services you connect still receive what each task needs.

To look before you leap, ./install.sh --check-only checks the machine without installing anything, and ./install.sh --dry-run generates the configuration without starting it. The manual installation described in the getting-started guide remains fully supported.
What does the guided installer do?
It asks its questions — in English or French — then does the whole job in one pass:
• generates a private configuration and the folders it needs
• builds the application, or fetches prebuilt images, and validates the settings before starting anything
• starts the stack: the database is migrated and the reference data — personalities, model catalogue and prices, default configuration — is loaded in a single transaction, all or nothing
• creates your administrator account and stores your provider keys encrypted
• restarts the server so every process uses them, runs a verifier, and prints a report containing no secret

❓ What it asks: how LIA will be reached, your administrator e-mail and name, the application's default language, an Anthropic and a DeepSeek key (the shipped configuration uses both; every model can be changed afterwards), an optional Google Gemini key (memory recall and searches by meaning run on it; the final report says what stays degraded without it), and whether to enable a few optional features — the monitoring stack, self-diagnostics, the script sandbox, the Live voice mode, the personal radio and the speaking avatar. Each stays off unless you answer yes. It also offers two optional pairs of keys, each given whole or left empty: a Google Maps browser key with its Cloud map ID, for interactive place maps, and — when LIA is reached on a public domain — a Cloudflare Turnstile site key and secret key, for an anti-robot check on sign-up, sign-in and password reset.

For an unattended install, the same answers can come from a file: ./install.sh --non-interactive --answers answers.env. To change these choices later — how LIA is reached, the optional features — run ./install.sh --reconfigure: each question shows the answer in force, which Enter keeps, and your data, keys, administrator and secrets stay as they are. Before anything changes, the settings your new answers would rewrite are listed with their current and new values, and nothing is applied unless you confirm; if the new configuration does not come back — or you interrupt it — the previous one is put back.
How will people reach my instance?
That is the one choice worth thinking about. The installer offers three:
• lan — home or office network, no domain name: the web interface is published on port 3000 of the machine, in plain HTTP
• proxy — you already run a reverse proxy (nginx, Traefik, Caddy) on the same machine: nothing new is published, and your proxy serves HTTPS
• caddy — you have domain names and want HTTPS handled for you: certificates are obtained and renewed automatically. Point the DNS records of your web and API domains at the machine before installing

In every mode, browsers talk only to the web interface; the API never gets a port of its own.

The mobile apps need an HTTPS address: choose proxy or caddy if you want them.
How do I know the installation really worked?
A server that answers is not yet a working installation, so the installer does not stop there. Once the stack is up, a verifier checks, without reading any secret:
• the database is on exactly the expected migration
• the reference data was applied in full, under the exact expected version
• your administrator account exists, is active and has administrator rights
• the stored provider keys can actually be decrypted
• every core role of the configuration points to a provider you gave a key for

Only then does the installer print its report: the address to open, your administrator account, how the instance is reached and where the database backups are written. Sign in there; Settings › Administration › LLM Configuration then lets you change any model or add an optional provider key.
Are my secrets safe during installation?
Yes, by construction.
• Your password and provider keys are typed at a hidden prompt, handed to the server on its standard input and stored encrypted in the database — never written to the configuration file, the installer's state or its log, and never on a command line
• Each required provider key is then sent once to its own provider, in a request header, to check it: a rejected key or an unreachable provider is reported at once and never stops the installation
• The server's own secrets — the encryption key, the session key, the database and cache passwords — are generated at random into a configuration file only your account can read
• The resume state keeps only non-secret facts and fingerprints
• An answers file for an unattended install is read once, never copied, and refused if other users can read it

Your administrator password must meet the instance's password policy; the installer states the rule when it asks.

One thing never to do: change the generated encryption key by hand. Everything encrypted with it would become unreadable.
What if a step fails halfway?
The installer stops at the first failing step, names it, writes the details to its own log with every secret removed, and gives you the command to continue:
./install.sh --resume

On a first installation, a failure only stops the containers it started: your configuration, data, backups and log stay in place. Resuming keeps your answers and the generated configuration, asks again only for the secrets it never stores — your administrator password and your provider keys, the optional one included — then runs the deployment steps again.

Resuming is deliberately cautious: if the files it generated changed in the meantime, it stops before touching anything rather than guessing. If you interrupt an installation yourself, the same command picks it up; an interrupted --reconfigure puts the previous configuration back instead — just run it again.
Prebuilt images or a local build?
The installer follows a rule, never a guess:
• from a clone of the repository, it builds the images from your source — the default, and the path that always works
• from an official release whose manifest is marked as qualified, it uses prebuilt images, each pinned by its digest — never a moving tag such as “latest”
• from a release whose manifest is missing or not qualified, it builds locally from the release's own verified source
./install.sh --local-build forces a local build in any case.

Limits of the prebuilt images: browser push notifications are not available (notifications arrive in the app), and changing a public setting that is built into the web interface requires a local build.
What does the installer not do?
It installs a working instance. A few things stay deliberate, separate acts:
• Upgrades: moving to a newer release is a written procedure (next question)
• Wiping and starting over: it never removes your data volumes; deleting data is always a command you type yourself
• Taking over an existing installation: if it finds a configuration it did not create, it refuses to touch it
• Changing provider keys: done later, signed in, in Settings › Administration › LLM Configuration
• Google and Microsoft: their sign-in and connectors are set up after installation, following the getting-started guide
• Searches by meaning (memory recall, knowledge spaces, the journal) run on a Google Gemini key: the installer asks for it but lets you skip it, and they stay unavailable until you add it in Settings › Administration › LLM Configuration

The installer itself runs on Linux only.
How do I move my installation to a newer release?
There is no upgrade command, on purpose: each step should be one you can see, stop and reverse. The self-hosting guide writes the procedure out (“Upgrading to a newer release”); its shape:
• back up first: a fresh database dump, your configuration and the current image pins — migrations run by themselves when the new version starts, and there is no way back down
• download the new release and check its SHA-256 checksum
• stop the application; your data stays
• copy the release's files over your installation, keeping your own configuration, overlay and installer state
• pin the images to the new release, start, and let the migrations run
• check the version and the services' health, then send one message

Rolling back the code means restoring the previous pins, in seconds; undoing a migration means restoring the dump — which is why it comes first.

Never set APPLY_SEEDS=true for an upgrade: the reference data deletes before it inserts. And ./install.sh --resume is not an upgrade path — it refuses once the release files change, by design. The procedure is written for an installation running a release's prebuilt images.
Can I run LIA without any cloud AI provider?
Partly. Here is exactly where the line is.

✅ On your own hardware: every model role — routing, planning, answering, agents — can run on a model served by your own Ollama server, set role by role in Settings › Administration › LLM Configuration. Dictation is transcribed on the server itself.

☁️ Still online:
• the guided installer asks for an Anthropic and a DeepSeek key, because the shipped configuration uses them; you can move every role to Ollama afterwards
• searches by meaning — memory recall, knowledge spaces, the journal — use Google's embedding service
• spoken answers come from online voice services, the default one free and without a key
• web search, weather, maps and the connectors you link are online services by nature

A fully offline installation is therefore not possible; one whose conversations are written by models on your own hardware is.
Can the mobile apps connect to my own server?
Yes: the official Android and iPhone apps work with any self-hosted LIA (see “Is there a mobile app for Android or iPhone?” in Getting Started). What your server needs:
• an HTTPS address: the apps refuse plain HTTP, so an installation reached in lan mode is for browsers only — choose proxy or caddy
• on iPhone, web and API addresses under the same domain (for example lia.example.org and api.example.org): Apple's browser engine blocks the sign-in cookie across different domains
• notifications: on Android, through a Firebase project you own and configure on the server; on iPhone, Apple lets only the app's publisher push, so your server must be given a wake relay's address — it carries one fixed sentence with no content, and without it the iPhone app receives no push
How do I give web search to accounts with no search key?
The guided installer does it by default: its web_search question is answered yes unless you refuse it, and it starts the searxng service that ships with LIA, with a generated secret. By hand: add searxng to COMPOSE_PROFILES, set SEARXNG_SECRET (for example openssl rand -hex 32) and SEARXNG_URL=http://searxng:8080, then restart. It asks public search engines (Google, Brave's site, DuckDuckGo, Bing News…) from your server's address, unfiltered, and serves every account that has no Brave key of its own. It is never exposed to the Internet; allow about 150 MB of memory.
How do I offer LIA's notifications without Google, with ntfy?
Run LIA's own ntfy server: the guided installer asks for it (ntfy_server, no by default) and starts the ntfy service that ships with LIA. Each person then subscribes in Settings › Push Notifications › ntfy notifications and receives LIA's notifications in the free ntfy app, on Android or iPhone, with no Google account. Without it, a person can still subscribe through a public server such as ntfy.sh. Behind your proxy or Caddy, the server needs a domain of its own (ntfy.example.org, for example); on a local network, phones reach it on port 2586. Only LIA publishes on it, with a token the installer generates; a topic's random name is what lets a phone read it. iPhones receive notifications at once only if you let your server wake them through ntfy.sh (the ntfy_ios_relay question), which then receives each message's id and a fingerprint of its topic's address, never its text; without it, an iPhone may receive them hours late. Allow under 20 MB of memory.

LIA is ready. And slightly impatient.

An assistant that acts, knows you and anticipates, while you stay in control of it.

Create my account