Appearance
Three Ways to Use AI
A kitloom site works with AI in three ways, and they answer different questions.
- AI Agents — your agent runs the site from the outside. Claude, Cursor or another MCP client connects to
/mcpand does what you ask it in its chat: writes and fixes posts, sets up SEO and redirects, looks at the day's orders. You talk to the agent; the site gives it tools. - Assistant — the same tools, in a chat inside the admin. A button in the corner of every admin page opens it, it knows which page you are on, and every change it proposes waits for your Approve. Nothing to install or connect: it answers with the site's own model, AI Writer's.
- AI Writer — the site uses AI itself, one text at a time. You save an API key on the site once, and the modules that write text — a translation, a meta description — ask the model when someone presses their "with AI" button. You press a button; the site asks the model.
Side by side
| AI Agents | Assistant | AI Writer | |
|---|---|---|---|
| Who starts it | You, in the agent's chat | Anyone allowed, in the admin's chat | Someone in the admin, with a module's button |
| Where it happens | In Claude, Cursor, claude.ai — outside the site | On every admin page, in a panel; and on Assistant → Chat | On the site, in the admin's screens |
| Whose model and key | The agent's: your Claude or Cursor subscription | The site's: AI Writer's key and model | The site's: an API key saved in AI Writer → Settings |
| Who pays | Nothing on the site's side; your subscription | The provider bills the key's owner, per question | The provider bills the key's owner, per question |
| What it can do | Everything its tools reach: content, media, settings, menus, SEO, shop, bookings… | The same tools as an agent | One text at a time, for the module that asked |
| What holds it back | The rights of the user whose key it holds; locked settings; nothing deleted for good | The rights of the user chatting; every change waits for Approve; locked settings; nothing deleted for good | A monthly budget; the asking module's own permissions; answers saved for a person to look at |
| What is written down | The activity log, under the key's user | The conversation; AI Writer → Usage with each answer's tokens and cost; the activity log | AI Writer → Usage: every question, its tokens and cost |
| Turned off by | AI agents → Settings, or revoking the key | The use_assistant permission, or the module switched off | Removing the key, or a spent budget |
| Packages | kitloom/ai + -filament / -nova | kitloom/assistant + -filament / -nova | kitloom/ai-writer + -filament / -nova |
AI Agents and AI Writer need nothing of each other: install one, both or none. The assistant needs both — AI Writer's model and the agents' tools — but no agent connected: /mcp may stay off.
Which one when
An agent is the better fit when you already work in Claude or Cursor, the work is long, and you would rather pay with your subscription than per question:
- "Go through last month's posts and give each a meta description."
- "Fill the site with sample content for the demo."
The assistant is the better fit for the rest of the team, inside the admin, with nothing to set up on their side:
- an editor on a post: "Shorten this excerpt", "Which category fits this best?";
- a shop manager on the orders: "Which orders from this week are still unpaid?";
- anyone who should see each change before it is made — the approvals are on unless the site owner turns them off, accepting the responsibility.
AI Writer is the better fit when the work is one step of an admin screen:
- an editor translates a post with one button;
- the site owner wants a hard limit on what AI may cost a month, and a record of every question;
- a module does it on its own, without a person in a chat.
The same task, three ways
A post's excerpt written by AI:
- With an agent: "Write an excerpt for the post 'Spring collection' and save it." The agent reads the post through its tools and saves the excerpt as the key's user would.
- With the assistant: on the post's edit page, ✦ Assistant → "Write a one-sentence excerpt for this post and save it." It reads the post it is looking at, shows the change in a card, and saves it once you press Approve.
- With AI Writer: only where a module offers a "with AI" button — today Translate with AI in Languages; there is none for the excerpt itself.
Safety in all three
- No more rights than a person. An agent acts as the user whose key it holds, the assistant as the user chatting; a module asks AI Writer only for a user it would let do the job by hand.
- The key stays where it is. An agent's key is an application password, revoked in one click. AI Writer's key, which the assistant uses too, is encrypted on the site and never shown again, not even to administrators.
- Every change can be undone. Each save keeps a revision of the record — its words, custom fields and SEO meta — and an agent's are marked as its own; an older version comes back from the record's Revisions in either admin, or with
restore_revision. - A text is only text. What a page or a post says never becomes an instruction: neither agents nor the assistant can change locked settings whatever they read — the assistant's own approvals among them — and AI Writer sends the instructions and the text apart.