These docs describe the current Dante AI platform. On the earlier platform? Your API reference is api-v2.dante-ai.com/openapi.json, and your key goes in the x-api-key header.
Sandbox
A private preview of your agent. Tweak its behaviour, swap models, and chat with it before any visitors see your changes.
When it opens
The Sandbox is on every plan, including Free. While the setup wizard is still running it is dimmed in the left menu as Test it privately, because there is nothing to test until your AI brain has something to learn. The line under it says so and links you to Add knowledge; once your first source has finished training, the row opens on its own.
What the Sandbox is for
The Sandbox is a split view: the Behavior and Model controls sit on the left, and a private chat preview runs on the right. Chatting and editing here are private until you press Save changes. Save changes applies your edits to your AI agent straight away, everywhere it is deployed, with no separate publish step. So it is the place to:
- Edit the agent's behaviour text and watch the chat react.
- Swap between AI models to see the cost and quality trade-off.
- Send test questions and see how the agent answers.
Replies stream in live, each word fading in smoothly as it arrives, exactly as your visitors will see them in the published widget. If your system settings ask for reduced motion, replies appear instantly instead.
One thing is deliberately different. When the Sandbox shows where an answer came from, it can name any of your knowledge sources, including uploaded files, text and Q&A answers, because this preview is private to you. Visitors on your live widget only ever see pages from your website named as a source. Your uploaded files, text and Q&A answers are never shown to a visitor as a source, even when your agent answers from them.
Behavior
The Behavior box is your agent's persona, the rules it follows on every reply: its tone, what to cover, and what to avoid. Write in plain English, "Speak warmly and keep replies to two sentences", "Always answer in Japanese", "Never discuss pricing", and the agent follows it literally.
Start from a preset (General AI agent, Customer support, Sales assistant, Product expert, or Lead qualification) and edit from there, or write your own from scratch. Use Reset to default to undo your edits, or the expand icon to edit in a full-screen window.
Refine instructions Starter rewrites a rough description into a polished agent prompt. It is available on paid plans, with 200 refines a month on Starter, 500 on Advanced, and no limit on Pro. See Refine instructions.
During onboarding, when the text still contains a highlighted token like [friendly], a Vibe picker appears above the box so you can drop in a starter tone (Friendly, Formal, Funny, Casual, and more) in one click.
During the first onboarding steps the Behavior controls start locked, and the lock badge tells you the exact action that unlocks them. On the naming step it reads "Name your AI agent first": add a name in the card above and hit Save changes. On the test step it reads "Test your agent first": chat with your agent in the preview on the right, and the moment you send your first message the Behavior controls unlock so you can fine-tune the persona based on what you saw. Clicking the locked area nudges the same instruction, so you are never left guessing.
While the vibe token is still in the text, the Behavior box itself is read-only; use the Vibe picker row above it to swap the token, then edit the text freely.
Model
Pick the AI model that powers replies. Models are grouped by tier, Fast and cost-efficient, Versatile and highly intelligent, and Most powerful at complex tasks, and each row shows how many credits it spends per message. Cheaper models cost fewer credits per message.
Free plans can use the two 1-credit models: GPT-4o Mini and GPT-5.4 Nano. Paid plans unlock the rest of the catalogue, including Gemini, Claude, and the larger GPT-5 family. Models above your plan show a lock icon, click one to see the upgrade options. During onboarding the locked rows are informational only: clicking one shows a short note and you can upgrade any time after setup.
The model picker unlocks as soon as your agent has at least one knowledge source, so you can start choosing a model while that source is still training. You do not have to wait for training to finish.
Want to weigh models against each other? Click Compare two models on the same question below the picker to open Side by side, where you can run one question through up to four models at once.
Save vs update
On a phone or tablet
On a wide screen the Sandbox shows your settings on the left and the live preview on the right, side by side. There is not room for both on a narrow screen, so below roughly 1024px wide the page shows one at a time with a Settings / Preview switcher at the top. It opens on Settings, because that is where Save changes lives, and one tap flips to the preview to see the result.
The left navigation also folds away on a narrow screen. Tap the menu button in the top left to open it, along with your workspace switcher.
Still stuck? We'll help.
Every plan, including Free, comes with human support. If something isn't behaving the way you expect after following the steps above, send us a note and we'll get back to you promptly.