The site assistant
Every site has an assistant that can answer questions about that site — its data, pages, forms and workflows. Open it from Assistant in the sidebar, or from the ✨ button that appears next to things you might have a question about, such as a workflow run that failed or a datastore table.
The ✨ button opens the assistant over whatever you are working on, with the question already written out. Nothing is sent until you press enter, and closing it puts you back exactly where you were — so you can ask something in the middle of editing a page or a workflow without losing your work.
What it can see
The assistant only knows what it looks up while answering you. It reads your site through the same tools an external AI client would use, and it can only reach the parts of the site you can reach: if you can read the datastore but not the pages, neither can it.
It is told never to make anything up about your site. If it has not looked something up, it should say so and then go and look — not guess at a plausible answer. If you ever see it describe a table or a page that does not exist, that is a bug worth reporting.
What it is allowed to change
Each conversation has a setting — Assistant can — with three choices. It always starts on Read.
- Read — look at the site and answer questions. It cannot change anything.
- Propose — also write page drafts and create workflows that stay switched off. Nothing goes live, and nothing runs, until you say so. Use this when you want to see the change before it happens.
- Apply — also change data, publish pages and run workflows.
Choices you do not have permission for are shown but cannot be selected. Raising the setting never grants you anything: every action is still checked against your own permissions on the site.
Why the setting exists. The assistant reads content other people wrote — form submissions, datastore rows, page text. Occasionally that content contains instructions aimed at the assistant itself, trying to get it to do something you did not ask for. The assistant is built to treat all of it as information to report rather than orders to follow, but keeping the destructive actions behind a separate, deliberate choice means a mistake cannot quietly change your site.
A good habit is to stay on Read for questions, switch to Propose when you want it to draft something, and only use Apply for a change you have already decided on.
Giving it standing instructions
If you are a site maintainer, Add instructions lets you record things the assistant cannot work out on its own — which table holds what, how your business counts something, which rows to ignore. These are given to it at the start of every conversation on the site, for everyone who uses it.
Good instructions are short and factual, for example: Revenue lives in the ledger table. Our financial year starts in April. Ignore rows where status is 'test'.
Things worth knowing
- Conversations are per site and private to you. Two people working on the same site do not see each other's questions, and a site's conversations never mix with other sites'.
- It uses the model your administrator has configured. Smaller models are faster but make more mistakes, including occasionally answering in the wrong language. If answers seem poor, mention the model to whoever runs your MFICore.
- Deleting a site deletes its conversations.
- Usage is billed to the site, both the tool calls and the model's own usage.
- If the Assistant item is missing, an administrator has not switched the feature on for this deployment.
Related
- The AI Agent — the general assistant on your account, which searches the web instead.
- AI access (connect your own assistant) — pointing Claude or another AI client at MFICore, which uses the same underlying tools.