AI
Manablox can work with AI models to help you write and build. Once a space has an AI provider connected:
- A magic wand appears next to text fields. It writes or rewrites the text for that field.
- Picture and video fields get a wand that generates an image or a video.
- The document editor gets a Generate button that writes the whole document at once, including its blocks.
- Content types, templates, workflows and even new spaces can be designed from a description in plain words.
Nothing of this is switched on by default. A space without a provider shows no wands at all.
You always stay in control: AI answers appear in a window first, you can change them, and nothing is saved until you save the document yourself.
Who can do what
Section titled “Who can do what”| Role | AI |
|---|---|
| Owner, admin | Connect providers, choose models, write the house style; and use everything below |
| Editor, author | Use the wands and Generate, and the Describe buttons on pages their role can open |
| Viewer | Nothing |
Connect a provider (admins)
Section titled “Connect a provider (admins)”An AI provider is the company (or your own server) that runs the model. Every generation is billed by that provider to the key you enter, so you need an account with them. Keys are set per space, so two spaces can use two different accounts.
| Provider | Does | Where you get a key |
|---|---|---|
| Claude | Text | An Anthropic API key, from console.anthropic.com. Starts with sk-ant-. |
| ChatGPT | Text | An OpenAI API key, from platform.openai.com. Starts with sk-. |
| Images and video | A Gemini API key, from aistudio.google.com. One key covers every Google model. | |
| Self-hosted | Whatever you set it up for | Only if your server asks for one. See Self-hosted models. |
For writing you need Claude, ChatGPT or a self-hosted text model. For pictures and video you need Google (or a self-hosted image model).
To connect one:
- Click Settings in the sidebar, then the AI tab.
- Under Providers, click the provider’s row, for example Claude. It opens.
- Paste your key into the key field, for example Anthropic API key.
- Leave the suggested model, or pick another under Model for text (or Model for image and Model for video for Google).
- Make sure Offer this provider in the editor is ticked.
- Click Save.
The row now shows “On”. Click Test the key: “Answered” means the key works. “No answer” shows the reason; usually the key is wrong or the account has no credit.
The key is stored encrypted. It is never shown again: the field stays empty on your next visit and only mentions the last four characters of the stored key. Leave it empty when you save other changes and the stored key is kept. Remove deletes the stored key; anything already generated stays where it is.
The house style
Section titled “The house style”Each provider has a System prompt: a few sentences sent along with every request in this space. Use it to describe your house style, for example “We are a family bakery in Vienna. Write warmly, in short sentences, and never use exclamation marks.” Leave it empty to use the default shown in the box.
Under What the dialog opens on you can preselect a writing style, tone and length (or, for Google, an image style, mood and so on), so editors start with the right settings.
Lengths
Section titled “Lengths”The Lengths card on the AI tab defines what “Short”, “Medium”, “Long” and “Extensive” mean in this space:
| Length | Words in a field | Words in a document | Blocks |
|---|---|---|---|
| Short | 30 | 250 | 3 |
| Medium | 90 | 700 | 5 |
| Long | 250 | 1500 | 8 |
| Extensive | 500 | 3000 | 12 |
You can rename them, change the numbers, or add your own. Back to the defaults restores the table above.
Self-hosted models
Section titled “Self-hosted models”If you run a model yourself (with Ollama, LM Studio, vLLM, LocalAI or a similar tool), click Add a self-hosted model on the AI tab. You fill in a Name, a Slug, what the server Speaks (OpenAI-compatible or Ollama), its Base URL, and What it can do (Write text, Draw pictures, with the model name for each). Test the connection checks that the server answers and has the models you named.
This usually needs a developer. A model running on the same computer or in your own network is blocked until it is in the instance’s allowed AI hosts. The developer adds its address to AI_ALLOWED_HOSTS in the project’s .env (for example localhost:11434 for Ollama) and restarts the CMS; see The .env file. The test tells you which address to add.
Writing a field with the wand
Section titled “Writing a field with the wand”- Open a document in Content.
- Next to the label of a text or rich text field, click the small wand. Its tooltip says, for example, “Write Summary with AI”, and a window with the same title opens.
- Choose a Writing style (for example Marketing or Technical), a Tone (for example Friendly) and a Length, if you like.
- In Prompt, say what the text should be about. Name the point to make, not the exact words: “Invite people to our summer bread course on 12 July, beginners welcome”.
- Click Generate (or press Cmd+Enter / Ctrl+Enter).
- The text appears under Result. Edit it right there if it is nearly right, or click Again for a new try.
- Click Use it. The text goes into the field.
- Save the document as usual.
The model is told where the text is going: the content type, the field’s name, the language you are editing, the document’s title and what the other text fields already say. A teaser written this way fits the body that is already there.
At the top of the window you can pick another Provider or Model for this one generation, for example a cheaper model for a first draft. The space settings stay as they are.
Generating pictures and videos
Section titled “Generating pictures and videos”Asset fields that take pictures get a Generate an image wand; fields that take videos get Generate a video. The Assets page has the same two buttons next to Upload files.
For an image you can choose a style, a Mood, an Aspect ratio and a Resolution (1K, 2K, 4K). For a video you choose a style, a Camera movement, an Aspect ratio and a Duration (4, 6 or 8 seconds). Then describe what should be in the picture or shot and click Generate.
A finished picture is saved to the asset library first, then put in the field. It is a normal asset: you can crop it, give it alt text and reuse it (see Images and files). It stays in the library even if you do not use it.
Videos can take a few minutes. You can close the window meanwhile; the video still lands in the library.
Generating a whole document
Section titled “Generating a whole document”Click Generate at the top of the document editor. The “Generate the whole document” window asks for a prompt and a length, and then writes the title and every text field of the document together, so the teaser and the body talk about the same thing.
If the content type has a block field, AI also picks which blocks the page needs, fills them in and arranges them on the grid. You see the result as a real preview of the block field: you can drag blocks around and change texts before anything reaches your document.
More controls lets you steer further:
- Blocks it may use: untick blocks the model should leave out.
- Describe pictures for the image fields: untick to get text only.
- Audience: who you write for, in your own words.
- Keywords: words the text should use, the most important first.
The pictures
Section titled “The pictures”When the document has image fields, the first answer contains a short description of each picture instead of the picture itself. If the space also has an image provider, a Draw N pictures button appears. The pictures are drawn one after another and appear in the preview as they arrive. You can also skip this and fill the image fields later with their own wands.
Click Use it to put everything into the document. It replaces what those fields held. Nothing is saved until you save, so you can still undo or discard.
Design with AI
Section titled “Design with AI”The same providers can build structure for you. Each design is shown in full first; nothing is created until you confirm.
| Where | You describe | You get |
|---|---|---|
| Content types > Describe a content model | What the website holds | Several content types and the block types they need |
| Content types > Describe it on a card | One document type or one block type | That type (plus blocks it needs) |
| Templates > Describe a template | What a template is for | A template with sample text |
| Workflows > New workflow > Describe it | When it starts and what should happen | A complete workflow, switched off |
| Settings > Spaces > new space > Describe it | What the new space will hold | A new space with its content types |
In the preview you can untick parts you do not want before you click create. A workflow design leaves values only you know (an email address, a web address) as placeholders and says so in its notes. When a design has problems AI could not fix, the preview lists them.
For a new space, you pick which existing space’s AI to use. Tick Give the new space the same AI providers to copy those providers, keys included, into the new space.
Designs work best with the strongest model you have.
Let an AI agent work here
Section titled “Let an AI agent work here”The Let an AI agent work here card on the AI tab is for the opposite direction: an AI assistant (such as Claude or ChatGPT in their own apps, or a coding assistant) that builds things in your CMS through its API. The card lists three addresses that describe your installation to such an agent, and a first message to paste into it. Give the agent an API key with only the permissions the job needs.
What happened, and what it costs
Section titled “What happened, and what it costs”Recent generations at the bottom of the AI tab lists the latest requests in this space: the prompt, the provider and model, and the reason when one failed. Every generation is also in the activity log, with who asked for it. Keys never appear in either.
Every click on Generate is one paid request to the provider. Nothing is generated in the background or cached. A self-hosted model costs whatever running it costs you.
When you move a space to another installation, the AI providers travel without their keys and arrive switched off. See Moving a space.