---
title: Use the integrated AI agent
description: Work with the assistant that ships inside the CMS, know what it can do, which model tiers you have, and how credits are counted.
updated: 2026-07-25
---
Every Quilium site comes with an assistant already connected to it. No connector to install, no token to copy, no [MCP](/en/docs/ai-mcp-skills) setup: it opens inside the CMS, it knows which site you are on, and it acts on that site with your account and your permissions.
Live examples. The requests below are the kind you send on the demo sites used throughout this documentation: Le Cinquième (a venue with a calendar), Midi (fully bilingual), Cairn (photo-heavy hiking guides) and Oscilla (a bilingual SaaS site).
## What it is
The assistant is a chat panel inside the CMS. It reads the site through the same operations an external agent would, and it writes through them too, so its work is ordinary Quilium data: pages, content blocks, collection entries, media, translations. Nothing it produces is a special AI object you would have to clean up later.
Three things come with it, and they matter more than the chat box itself:
- **Site context, injected.** It starts each conversation knowing the site, your account, your profile on that site, and the site's custom instructions.
- **The [skills library](/en/docs/skills), on demand.** It loads Quilium's canonical patterns before acting, rather than guessing at YAML shapes.
- **Persistent conversations.** Threads are saved per site, so you can come back to one and it still knows what you decided.
Placeholder: screenshot of the assistant panel open inside the CMS, showing the thread list, the model picker and the usage indicator. To capture once the new CMS build is out.
## What you can ask it
Its centre of gravity is running the site day to day. On Le Cinquième, adding a concert is one sentence:
```prompt
Add the concert of 12 October, doors at 8pm, in the events collection.
Title "Nuit blanche", room "Grande salle", ticket link on the site. Show it to me before saving.
```
The same applies to the rest of the editorial routine: creating and filling pages, dropping content blocks, reorganising a menu, uploading and sorting media, writing meta titles and descriptions, auditing what is missing for SEO.
**Translations.** On Midi, "translate the three new drinks into English, keep the French as the source" is a normal request. It writes one locale per call, the way the platform requires.
**Media.** On Cairn, it can describe a batch of photos and fill their alt text, then move them into the right folder. Uploads keep the original file, image versions stay defined by the content-type.
**Beyond the CMS.** It reads the files you drop into the conversation (PDF, spreadsheets, images), searches the web when you ask for a reference, generates or edits images, and writes the HTML and CSS of an embed block. On Oscilla, "here is the release note as a PDF, draft the changelog entry from it in both locales" is a single request.
**Elsewhere in the CMS.** The same engine powers the AI actions outside the chat: generating a description for a media item, translating a field, generating or editing an image, dictating text instead of typing it. Those actions draw on the same credits.
## Which model you get
The picker offers three tiers. They differ in speed, cost and how much reasoning they can sustain.
| Tier | Use it for |
|---|---|
| **Éclair** | Short answers, classifications, repetitive batches. The cheapest and fastest. |
| **Standard** | The default. Most editorial and creative work. |
| **Premium** | Long-form writing, complex HTML and CSS, deep reasoning on a tricky structure. |
Which tiers a site may use is set on the site itself, so a picker that shows a locked tier is telling you the plan does not include it, not that the model is down. Premium is opt-in. When a thread asks for a tier the site cannot use, the request falls back to the strongest one it can.
Placeholder: which plan includes which tiers, and where an admin changes it. To fill from the commercial offer once it is final.
## How credits are counted
Every billable action writes a usage record: a chat turn, a delegated sub-task, a media description, a translation, an image generation or edit, a voice transcription. Each one costs **credits**, derived from what the action actually consumed. The consequences are the ones you would expect: a long conversation costs more than a short one, Premium burns credits faster than Éclair, and generating a high-quality image costs more than describing one.
Credits are counted **per site**, over three calendar windows at once:
| Window | Resets |
|---|---|
| Daily | At midnight |
| Weekly | Monday |
| Monthly | The first of the month |
Reaching **any** of the three stops new requests until that window resets; the panel says which one it is and when it comes back. Nothing you already produced is affected, and reading the site never costs anything. Cairn, for instance, runs on 50 credits a day, 100 a week and 150 a month: enough for a full editorial day, not enough to regenerate the whole site every morning.
Credits are shared by the site, not by the person. Everyone working on the same site draws from the same daily allowance. On a busy day, keep Premium for the tasks that need it.
Placeholder: the actual credit allowances per plan, and the wording of the limit banner. To fill from the commercial offer once it is final.
## What it will not do
The guardrails are deliberate, and they are the same ones that apply to [any agent](/en/docs/ai-mcp-skills) operating a Quilium site.
- **It does not publish pages.** Page work lands unpublished, and going live is a human action, page by page. Collection entries are the exception to know: they have no draft state, so a saved entry is live wherever a query picks it up.
- **It confirms before destroying.** Deleting, unpublishing or overwriting existing content is announced and confirmed first.
- **It stays inside your profile.** Asked for something your profile does not allow, typically a structural change on a content-only profile, it says so instead of half-doing it.
## Steer it once instead of repeating yourself
A site carries **custom instructions**: a short text injected into every conversation on that site. It is the right place for what never changes, and it saves you a paragraph in each prompt.
```text
Editorial tone: informal, second person, short sentences.
Prices always in euros, format 12,50 €.
Never touch the "legal" pages.
```
Placeholder: menu path where an admin edits the site's AI custom instructions. To capture with the new CMS build.
## When to use your own agent instead
The integrated assistant is built for running an existing site. It is not the right tool for building one from a prototype, where the work happens in your editor, in your repository, alongside the [templates](/en/docs/build-a-template) and the CLI.
| Situation | Use |
|---|---|
| Publishing, translating, media, SEO, day-to-day edits | The integrated assistant |
| Building the site, writing Liquid, wiring routes, bulk migrations | Your own agent, [connected over MCP](/en/docs/connect-agent-mcp) |
Both drive the same site through the same operations, so switching between them costs nothing.
## Related