Turn on the built-in AI assistant
Contents
- What does the built-in AI assistant do?
- Turn on the built-in AI assistant in 5 minutes
- Which model should you choose, and what does it cost?
- Add an image generator
- Write the editorial profile
- Budget, quota and editor access
- The cards: what the assistant did
- The [To complete: …] placeholders
- Security and privacy of your keys
- If the assistant doesn't respond
- Key points
- Going further
- Frequently asked questions
Éclair's built-in AI assistant writes drafts, improves your texts and creates images in your site's voice, without leaving the admin. It works with Claude or ChatGPT and your own API key: no Éclair subscription, you pay the provider for what you actually use. This lesson takes you from creating the key to your first draft, security settings included.
What you'll be able to do
Turn on the assistant
Create your key, test it and choose the right model in 5 minutes.
Make it write like you
Write your editorial profile and set a monthly budget.
Stay in control
Read the cards, undo a change and protect your keys.
What does the built-in AI assistant do?
You talk to it in plain language, as you would to a colleague: “Write an article about servicing an e-bike in spring, with an image.” It first checks that the topic doesn't already exist on your site, creates the image, writes the complete draft (title, Key takeaways, subheadings, lists or tables, FAQ, meta description, internal links), then shows it to you with its SEO and GEO scores. Nothing is published without your click.
You'll find it in three places:
- in the menu, AI assistant (shortcut
Ctrl+J) - on the dashboard, in the What do you want to create today? field
- in the editor, to touch up a passage or the content you have open
With the built-in AI assistant, your API key does the work: Éclair calls the provider for you. With the MCP connector, it's the other way round: you stay in Claude or ChatGPT, with your subscription, and they come to read your site. The two can work side by side.
Turn on the built-in AI assistant in 5 minutes
- Create an API key with the provider of your choice: console.anthropic.com for Claude, platform.openai.com for ChatGPT. Add a few dollars of credit to that account.
- In Settings › AI assistant, choose the Provider: Claude (marked “Recommended”) or ChatGPT. The Off option means no data is sent to any AI.
- Paste your key into the Claude (Anthropic) API key or ChatGPT (OpenAI) API key field.
- Choose the Model: each line in the list shows its estimated cost per article.
- Click Test the connection. The test is free: Éclair only reads the model's details, without generating any text.
- Click Save.
The assistant appears right away in the menu, in the editor and on the dashboard.
Which model should you choose, and what does it cost?
You pay the provider for what you actually use. Éclair calculates each expense from the billed tokens and shows it in Settings › AI assistant. For a full post, about 25,000 input tokens and 7,000 output tokens (searches and checks included), here are the official guide's estimates:
| Model | Label shown | Price per 1 million tokens (input / output) | Full post |
|---|---|---|---|
| Claude Opus 5.5 | the most capable | $4 / $20 | ≈ $0.24 |
| Claude Sonnet 5.5 | balanced | $2 / $10 | ≈ $0.12 |
| Claude Haiku 5.5 | fast and affordable | $0.10 / $0.50 | < $0.01 |
| GPT-6 Astra | the most capable | $10 / $50 | ≈ $0.60 |
| GPT-6.1 Sol | balanced | $2 / $10 | ≈ $0.12 |
| GPT-6.1 Terra | fast | $1 / $5 | ≈ $0.06 |
| GPT-6 Luna | affordable | $0.10 / $0.50 | < $0.01 |
Public prices taken on October 29, 2026 from the official pages of Anthropic and OpenAI; they may change. The assistant's fixed instructions are cached by the provider, which lowers the cost of follow-up requests.
To get started, Claude Sonnet 5.5 offers a good balance between quality and cost. You can switch models at any time in the settings.
Add an image generator
In the Images section, the Image generator field offers four choices. Generated images are optimized as WebP and stored in the media library.
| Choice | Strengths | Key needed | Indicative price shown |
|---|---|---|---|
| None | The assistant uses your media library and the photos you attach | none | $0 |
| GPT Image 2.5 | Versatile, follows instructions very well | your ChatGPT key, or a dedicated OpenAI key | ≈ $0.05 per image |
| FLUX 3 | Very realistic photos | Black Forest Labs key (dashboard.bfl.ai) | ≈ $0.05 per image |
| Ideogram 3 | Visuals with text (posters, banners) | Ideogram key (ideogram.ai) | ≈ $0.06 per image |
One rule protects your customers: for a product page, the assistant never generates an image of the product itself, because it would mislead buyers. It uses your real photos.
Write the editorial profile
This is what makes the AI write like you. The profile has six fields of up to 2,000 characters each: Who you are and what you offer, Who you write for, Tone and style, Values, Rules to follow and Image style.
- Click Draft my profile from my site: the AI reads your pages and posts for 20 to 60 seconds.
- Review the result and correct it. Add concrete rules, for example “always link to the contact page” or “never mention prices”.
- Click Save.
Budget, quota and editor access
The Limits and privacy section helps you avoid unpleasant surprises:
- Maximum monthly budget (dollars): $10 by default, 0 for no limit. It's checked before each step of a request; once it's reached, the assistant stops until the following month. A request in progress can go over it by one step at most.
- Requests per person per day: 60 by default, 0 for no limit. Each person can run at most two requests at a time.
- Editors can use the assistant: checked by default. Administrators always have access.
Below the budget, the “This month” line shows the amount spent and the number of requests and images, with a progress bar.
The cards: what the assistant did
Each action produces a card below the reply:
| Card | What happened | Your buttons |
|---|---|---|
| Draft | New post or new page, never published | Open, Preview, Publish now or schedule |
| Suggested change | For published content: nothing changes until you approve | Preview, Apply, Dismiss, then Cancel if needed |
| Updated | A draft changed directly | Open, Preview, Cancel |
| Changed in the editor | The content you have open is changed in the editor | Cancel, then you save |
| Image | Image created, optimized as WebP and stored in the media library | Featured image (in the editor) |
“Cancel” works as long as nobody has edited the content since. Otherwise, the earlier versions remain in the editor's Version history, which keeps the last 20. The assistant keeps your last 30 conversations.
The [To complete: …] placeholders
The assistant never makes up facts about your business: prices, opening hours, figures, testimonials. When information is missing, it leaves a visible placeholder such as [To complete: price] and points it out to you. The Analysis tab counts them as long as any remain: fill them all in before publishing.
Security and privacy of your keys
Your keys are encrypted on your server (libsodium, or AES-256-GCM). They are never sent to the browser or displayed: only a hint such as “sk-ant…ab12” appears. They are excluded from backups, as is the conversation history. For stronger protection, add this line to eclair-config.php before saving your keys:
define('ECLAIR_SECRET', 'a-long-secret-phrase');The other safeguards to know about:
- What goes to the provider: your request, attached images, the editorial profile and the relevant context (the list of your pages and posts, the content open in the editor). Don't include sensitive data.
- Data use: according to their terms, checked on October 29, 2026, Anthropic doesn't train its models on data sent through the API, and neither does OpenAI by default.
- Limited powers: the AI can't publish, delete, or touch the settings, menus or accounts.
- Prompt injection: content the AI reads is treated as data, never as instructions.
If the assistant doesn't respond
| Message or symptom | What to do |
|---|---|
| “API key rejected by…” | The key is incomplete or revoked: paste it again, then Test the connection. |
| “Out of credit at…” | Top up your account with the provider, then try again. |
| “This month's AI budget is used up…” | Raise the budget in Settings › AI assistant, or wait until next month. |
| “You've reached your limit of… assistant requests” | The daily quota is used up: raise it or wait until tomorrow. |
| “Model not available with this key…” | Choose another model in the settings. |
| “Can't reach…” | Your host blocks outgoing connections: ask them to allow api.anthropic.com or api.openai.com, plus api.bfl.ai or api.ideogram.ai for images. |
| The reply arrives all at once | A buffer at your host is holding back the stream. Everything works; on Nginx, add fastcgi_buffering off; (see Nginx, VPS, proxy and CDN). |
| The assistant doesn't appear | It needs the cURL extension and encryption (sodium or openssl), available at almost every host. |
Key points
- The built-in AI assistant works with your Claude or ChatGPT API key: about $0.12 per article with Claude Sonnet 5.5 (prices as of October 29, 2026).
- Test the connection is free: run it after each new key.
- The editorial profile and Éclair's rules make the AI write in your voice, with high SEO and GEO scores from the very first draft.
- The monthly budget ($10 by default) and the daily quota (60 requests) keep spending in check.
- The AI prepares, you publish: drafts and suggestions only, with keys encrypted and excluded from backups.
Going further
- Create a page with AI and polish its text: the Create with AI box and touching up a passage.
- Understanding the MCP connector: work from Claude or ChatGPT with your subscription.
- GEO: getting cited by AI: why Key takeaways and the FAQ matter so much.
- Troubleshooting: other errors and how to fix them.
- Official sources: Claude API pricing and OpenAI API pricing.
Frequently asked questions
Do I need a Claude or ChatGPT subscription for the built-in assistant?
No. The assistant uses an API key, billed by usage on your developer account (console.anthropic.com or platform.openai.com). To use a subscription you already have, go through the MCP connector.
How much does a full article cost?
About $0.12 with Claude Sonnet 5.5 or GPT-6.1 Sol, and $0.24 with Claude Opus 5.5, based on prices checked on October 29, 2026. An image adds about $0.05.
Can an editor use up the whole budget?
No. The monthly budget applies to the whole site, and the daily quota to each person (60 requests by default). You can also uncheck “Editors can use the assistant”.
Are my API keys included in backups?
No. They are encrypted on your server and excluded from backups, as is the conversation history. After restoring to another server, paste them again in Settings › AI assistant.