The AI assistant

What it does

The assistant has its own page in the dashboard, AI Assistant. You ask in plain words, in English or Arabic, and it answers from the dashboard's own data:

• Figures and cards. User, project and task figures as the dashboard's own stat cards; a split (verified, unverified and deleted users, projects by status, open and done tasks) as a donut; records as a list or as the dashboard's table; one user or project as a profile card. Each card links to the screen it comes from. A list or a table shows at most 5 rows, says "Showing 5 of" the total, and ends with a Show more link to the page holding the rest.
• Charts. New users, projects created or your own tasks added, per day or per month, as bars, a line, an area, or bars stacked by status, in the colour you ask for.
• Lookups. Find a user by name, email or username; list the newest sign-ups, the deleted users, the admins with their roles, the projects (by name, status or environment, or the deleted ones) and your own tasks; read the app settings as a table.
• Changes. Create, edit, delete and restore users, and change an app setting.
• Images. Attach up to 6 images of up to 10 MB each, which the model reads.

Conversations are saved per admin and listed in a history sidebar, where they can be renamed, archived and deleted. The empty screen suggests questions about what your data holds right now, such as the unverified users or this week's sign-ups, and each answer can end with follow-up suggestions. Where the model thinks before answering, its reasoning shows in a panel you can open.

Each set of data is shown once: as a card, or, when you ask for CSV or another format no card gives, as text with no card. A table of records is always the dashboard's table card, never a table written in the reply. A table the reply does write, such as two periods side by side, has buttons to copy it and to download it.

Providers and who pays

• Each provider runs on a key the API holds, set by whoever installs the dashboard. Every message is billed to that key, at the provider's prices, unless the person has added their own key for that provider: their own key is always used first.
• A provider with no key on the server still works: the assistant asks each person for their own key for it. That key is kept in their browser only, sent with their own requests, and never saved or logged on the server.
• One key is enough: the picker lists every provider's models, and a model whose provider has no key asks for one when it is picked.

What it is allowed to do

• Using the assistant needs the ai_chat.use permission. Of the seeded roles, Super Admin, Admin, Manager and Editor have it; Viewer does not.
• Every tool the assistant runs checks the signed-in admin's own permissions, the same ones the dashboard's routes check. An admin who may not delete users cannot have the assistant delete one, and the model is not even told about tools the admin may not use.
• Projects need projects.view, as on the Projects page. Tasks are always the signed-in admin's own quick tasks, so every admin may ask about theirs and nobody else's. A chart checks the permission of the series it draws.
• A deleted user is soft deleted, as on the Users page, and can be restored.

Without keys

• With no key on the server, the assistant opens the key dialog: "The server has no key for this model's provider. Add your own API key to keep going."
• A key whose quota is used up answers that the model has no requests left, and suggests another model.
• On the built-in mock API, the assistant offers one model, Aniq Demo, and answers from the sample data with scripted replies, cards and charts. Ask it "Show me every card" to see each kind at once. It needs no provider and no key.

• The built-in mock API

Keys and models

Tools, prompts and widgets

The MCP server