AI assistants

You can connect an AI assistant to LibreTimes and let it work on your account: search the index, read a work, check a bibliography against real records, keep your citation library tidy, and draft new work.

The connection uses the Model Context Protocol (MCP), which is the standard way an assistant talks to an outside service. You do not install anything – LibreTimes runs the server, your assistant connects to it over the web, and you sign in with your ordinary LibreTimes account.

• Connecting an assistant — the setup, per client.
• What an assistant can do — the full list, and what each one needs permission for.

Ask AI on a work

Separately from connecting your own assistant, work pages carry an Ask AI button when the feature is switched on. It opens a panel where you can ask about the text you have open — a definition, a proof step, what a section is claiming — and the answer is drawn from that one work and nothing else. It cannot browse, act on your account, or reach any other page.

You need to be signed in to ask, there is a limit on how many questions you can ask per hour, and every answer is labelled with the model that wrote it. Your conversations are kept for you to come back to and are private; you can delete them from the panel. If you would rather not see the assistant at all, turn it off in Settings → Privacy and the page is just the page.

AI in the editor

When the feature is on, selecting text in the editor shows an AI button. Its menu works on the selected passage only: improve the writing, shorten, expand, fix spelling and grammar, translate into another of the site's languages, explain it for a student, or give your own instruction. The result appears in the text as a suggestion: removed words struck through, new words highlighted, a ✓ / ✕ on each change and an Accept all / Reject all bar under the text. Nothing is final until you accept — rejected text goes back exactly as it was, and closing the editor with a suggestion still open rejects it. You can keep writing elsewhere in the draft meanwhile.

Only the passage you selected is sent to the model — not the rest of the draft. Accepting an edit marks the work as AI-assisted in its disclosure field; you can say more than that, but not less. The same per-hour limit as Ask AI applies, and the same switch in Settings → Privacy turns it off.

The Assistant button in the editor's header opens a chat beside your draft (a panel on a wide screen, a sheet on a small one). It is a writing assistant that does not read the draft: ask it anything, and when you want it to look at a passage, select the text and press Add selection to attach it to your message. Every answer has an Insert at cursor button — Replace selection when text is selected — which puts the answer into the text as the same kind of suggestion, for you to accept or reject there; accepting marks the work as AI-assisted like an accepted edit does. Conversations are kept per draft, so you can come back to them; the history menu lists them and lets you delete one.

What is sent to the model provider, and to whom, is set out in the privacy policy.

The one rule worth knowing first

An assistant drafts. You publish.

There is no tool that makes anything public. An assistant can create a draft, rewrite a draft, build a private course and add lectures to it, and fill in your CV — but the act of publishing is yours, taken deliberately, in the web app. Nothing an assistant does appears on your public profile, in search, or in anyone's feed until you put it there.

This is deliberate and it is not a temporary limitation. An assistant works fast and does not always work correctly, and a mistake that reaches your public academic record is not the kind of mistake you want to find out about later. The same reasoning is why our content policy treats undisclosed machine-written work as a problem: the value of a profile is that a person stands behind it.

What an assistant cannot do

Some things are left out on purpose, and they will stay out.

What it can see

An assistant sees what you would see signed in as yourself, and nothing more. Public works, public profiles, your own drafts, your own library and citations, your own CV. Other people's private material is as invisible to it as it is to you.

A word about prompt injection

An assistant reads text that other people wrote — that is the point of it. Text can contain instructions, and an assistant can be fooled by them. This is a real and unsolved problem across the whole industry, not something particular to us.

Our answer is to keep the blast radius small: the tools that write are limited to your own drafts, your own library and your own CV, nothing an assistant does becomes public on its own, and every tool that could reach another person was left out. Read what an assistant produces before you publish it, the same way you would read a draft from a co-author.

Permissions

When you connect, you are asked to approve what the assistant may do. The read tools need no permission at all. Each kind of write is a separate permission, and you can grant some and refuse others — an assistant allowed to manage your citation library is not thereby allowed to touch your CV.

The tool reference names the permission each tool needs.