AI and the Assistant

tilspire can use an AI model to help you run your site: ask about it in plain language, see what needs your attention, and have pages and posts drafted for you to approve. It arrived in 1.13.0, and since 1.14.0 the panel switches into an AI mode for it. You bring your own key, you set what it may spend, and nothing it writes goes live without you.

Connecting a provider #

1. As an administrator, open System ▸ AI settings and add a provider — Anthropic, OpenAI, Gemini or OpenRouter — with an API key from that provider. The key is encrypted on your server and never shown again; the page shows only its last four characters.
2. Press Test connection to check the key works.
3. Set a monthly spending cap. There is no default: until you set one, every AI call is refused, so nothing costs money before you choose a limit.
4. Choose a model for each kind of task, such as Ask for the chat and Digest for summaries. You can also offer up to 12 models that people may switch between in the chat.

Adding or replacing a key, turning a provider on, and changing the cap or a price ask for your password and a code from your authenticator app. Turning a provider off never does, so a leaked key can be stopped at once.

The Assistant #

Enable the Assistant module under System ▸ Modules. Ask Tilspire then appears at the foot of the admin menu: it, the AI switch in the top bar, or Ctrl+J (⌘+J on a Mac) turns the panel into AI mode, and Esc takes you back to the page you left. Ask it about your settings, your site's status, your pages and posts, menus, templates or forum activity. It reads only what you yourself may read, and it can change nothing on its own.

Since 1.17.0, everything AI in the panel needs the Use AI features permission, so a role can be given no AI at all. It replaced Use the Assistant: roles that had that keep AI, and on a site that never used the Assistant, grant it to your editors under Roles for Draft with AI.

• It proposes, you decide. Ask for a new page or post, or a change to one, and it shows the proposal as a card with every change, field by field. Approve it and it is saved as if you had made it in the editor; reject it and nothing changes. Undecided proposals also wait in the Review queue. Since 1.16.0 it can propose a new Layout, or a change to a Layout or a field, for anyone allowed to manage templates.
• You see it work. Each step appears as it happens: what it looked up and what it found. Stop ends it between steps, and always before it proposes anything. With Anthropic, OpenAI and Gemini the answer appears word by word; OpenRouter answers whole.
• What needs your attention lists tickets waiting on you, forum topics nobody answered, drafts to review, a waiting update, staff without two-factor sign-in, API keys about to expire and AI spending near the cap — each only if your role can act on it. Summarise turns it into a short, prioritised note.
• What your members write is treated with care. Forum posts and similar text are read as data, never as instructions, and once the assistant has read them it cannot propose anything in the same question. A post that tries to instruct an AI is flagged.
• Email addresses are masked before anything reaches your provider, unless you choose otherwise under System ▸ AI settings.
• Chats are saved since 1.15.0, readable only by the person who had them, and deleted after the retention period you set under System ▸ AI settings, or never kept if you switch saving off. Anyone can export or delete their own. Every step the Assistant takes is in the activity log, without the content.

More it can do #

• Where the panel opens: with the Assistant set up, the admin panel opens on its chat. See The admin panel.
• Email digests of what needs your attention, from your Profile.
• Forum moderation, if you switch it on per kind under System ▸ AI settings: it judges new topics and replies, and hides or holds a bad one for a moderator. Its decisions can be reversed and are logged, and the author sees a short notice.
• Translations drafted for you to approve. See Languages.
• Public answers on your search page, with their own budget. See Site search.
• Design proposals: colours, fonts and layouts for the active theme, previewed before you decide. See Themes.

What it costs, and how well it does #

Every call is metered: which model, how long it took, the tokens it used and what it cost. System ▸ AI settings shows the last 30 days per task and model — calls, failures, speed, cost and how many proposals you approved — and suggests a cheaper, faster or more reliable model only when a pattern persists, always with its numbers. None of this leaves your site.