The AI Agent

The AI Agent is a chat assistant built into your dashboard. It runs on a language model your cluster hosts itself, and — when your administrator has enabled web search — it can look things up on the web and answer with links to where it found them.

Open it from AI Agent in the sidebar, or press Cmd+J (Mac) / Ctrl+J (Windows and Linux) from anywhere in the dashboard. On the mobile app it is in the menu, beside Mailbox.

Asking a question

Type your question and press Enter. Use Shift+Enter to add a line break without sending.

The assistant decides for itself whether it needs to search. Ask it something timeless — "explain what a DNS CNAME record does" — and it answers straight away. Ask it something current — "what happened in the news today?" — and it will search the web first, read the most useful result, and then answer.

Watching it work

While the assistant is working you will see what it is doing and how long it has been doing it:

• Thinking — deciding how to answer.
• Searching the web — the search terms it chose are shown next to the label.
• Reading — the site it is currently reading.
• Writing the answer — producing the reply you see appear.

A timer shows how long the current step and the whole answer have taken. Long research questions can genuinely take a minute or more; if one is taking a while, the assistant says so rather than leaving you watching a number.

You can press Stop at any time. Whatever it had written is kept and marked as stopped, so nothing you were reading disappears.

When it finishes, a short summary appears under the answer — for example Answered in 24s · 2 searches · 1 page read — so you can see how much work went into it.

Sources

When an answer is based on pages the assistant read, those pages are listed underneath as Sources. Click one to open it. If an answer has no sources, it came from the model's own knowledge rather than from anything it checked just now.

This distinction matters. A model's knowledge has a cut-off date and can be out of date or simply wrong. If accuracy matters to you, prefer answers that cite sources, and follow the links.

Conversations

Each conversation keeps its own history, so you can refer back to earlier answers without repeating yourself. Use New chat to start fresh — a new conversation is the right choice when you change topic, because everything in a conversation is context for what comes next.

Conversations are titled automatically from your first message. You can rename or delete any of them, and they follow your account, so you will see the same history whichever device you sign in from.

What it cannot do

• It cannot see your MFICore data. It has no access to your sites, datastore, forms or workflows. Ask it how to do something in MFICore and it can help; ask it what is in your datastore and it cannot look.
• It cannot reach anything private. Web pages only, and only public ones — it is blocked from internal and private addresses.
• It can still be wrong. Everything a language model produces should be checked when it matters, especially numbers, dates, quotations and anything you would act on.

If something is not working

• "The AI Agent is not enabled" — your administrator has not turned it on for this server. They can do so under Infrastructure, in the admin dashboard.
• "Web search is unavailable" — either no search service is configured, or your administrator has switched web access off. Either way the assistant answers only from what the model already knows, and answers may be out of date.
• "You already have a message in progress" — one question at a time. Wait for the current answer, or press Stop.
• "You have reached the daily limit" — there is a cap on how many questions each person can ask per day. It resets automatically.
• Answers ignore what was searched — tell your administrator; this usually means the model's context window is set too small for the pages being read.