Using AI chat from the browser

The chat window is in your browser; the model answering it is not. It runs on the device you picked, using that machine's processor and memory. Nothing you type is sent to an AI provider, because there is no AI provider in the picture.

Before it will answer

Three things have to be true:

  1. A device is selected and its agent is online.
  2. The AI service is running on that device — it is separate from the agent and does not start with it unless AiAutoStart is on.
  3. A model is loaded. If none is, the chat says the agent is not loaded rather than pretending to think.

Starting the AI service and choosing a model can both be done from the browser — see AI models and the knowledge base. If the service will not come up at all, work through the AI does not load.

The monthly message count

Above the input you see how many AI messages are left this month, and when they run out the chat says the monthly limit is reached. The allowance comes with your plan: in the shipped defaults even the free plan gets a modest monthly number and the top plans are unmetered. A plan set to zero shows an "upgrade required" screen instead of a chat. The numbers are in plans and quotas.

Only messages you send through the website count. The model runs on your own hardware, so this is a service limit, not a compute one — and asking the same question in the desktop app draws on the same monthly allowance.

Sessions

Each conversation is a session, stored on the device, not on the server.

  • New session starts a fresh one.
  • Rename gives it a name you will recognise later.
  • Export downloads the conversation.
  • Clear empties the messages but keeps the session.
  • Delete removes it permanently, with a confirmation. There is no undo.
  • Summarise asks the model to condense the conversation so far into a few bullet points, and posts the result as a message. The request itself is not stored as part of the chat.

Sessions started on the desktop app and in the browser are the same sessions — both talk to the same AI service.

The knowledge base chip

Next to the session controls, a chip reports what the knowledge base is doing: how many documents are indexed, whether retrieval is enabled, or off when it is not in use. In the browser this chip is a read-out; you change what is indexed from the AI management screen.

When the AI service is stopped, the chip says so plainly instead of showing stale numbers — and it can still show what is on disk, marked as coming from disk rather than from a running service.

Letting the AI touch your files

Allow file management lets the model create, rename, move and delete files — inside your shared folders only, never outside them. Every change asks you to confirm before it happens — creating, renaming, moving and deleting alike — and the actions it took are listed under its reply.

Two things to keep in mind:

  • The limits are the same ones the file manager obeys: whatever is not in rootPaths does not exist as far as the AI is concerned. See what the agent may reach.
  • A model is not a careful colleague. The switch is off until you turn it on (since version 1.11.5) — but a machine that had it on before an update keeps it on, so check where it stands before pointing the AI at a folder of irreplaceable files.

Searching the web

The AI does not reach the internet on its own. Web search is off until you switch it on, and while it is off nothing about your conversation leaves the device.

When you do switch it on, a search is not sent silently. The AI first shortens your question into a search string, strips out what identifies you — names, companies, document titles, addresses, long numbers — and then shows you the exact text before it is sent. You can edit that text, send it as it stands, or decline: declining still answers the question, using only what is on your machine.

What we cannot hide is the topic. The search engine sees the words you approved and roughly what you were asking about, along with the fact that a search happened at all. Content stays private; subject matter does not. If that matters for a particular question, decline the search — the answer will come from your own files and the model's own knowledge.

Everything that does go out is recorded on the device. In the AI settings, under Web search, "Show what left this machine" lists each outbound search with its time, its destination and the exact string. The record covers what was sent — never what came back.

When answers are slow or strange

  • Slow first answer, fast afterwards is the model being loaded into memory.
  • Empty replies usually mean the model ran out of room for its answer; the reply-length cap is AiMaxNewTokens in the agent settings.
  • Answers that ignore your documents mean retrieval is off or nothing is indexed — the chip tells you which.

Updated 2026-10-16

This didn’t help — contact support