These docs describe the current Dante AI platform. On the earlier platform? Your API reference is at api-v2.dante-ai.com/docs (sign-in needed), and your key goes in the x-api-key header.
Quickstart
Your AI agent answering a real question with your own content in a few minutes, then embedded on your site.
1. Create your agent
There are two ways in. If you don't have an account yet, enter your website address on the homepage and press Show me. Dante reads your site right there, page by page, and lets you ask your new AI agent a real question about your own business before you create an account. When you decide to keep it, signing up keeps the very agent you just watched being built, so your site is never read twice and nothing you did is lost. Once you have signed up you land on your dashboard, which picks up that agent's setup at the step you had reached, or simply shows your dashboard if setup is already done. If the preview is not available, you go straight to the Knowledge step instead and create your account there, with your address carried over. If you close the create-account card without signing up, you go back to the homepage you came from with your website address still in the box, rather than to a second screen asking for it again. If you already have an account, open the dashboard, click New AI agent, and give it a name in the create dialog (you can change it any time).
You can also sign in from the setup screen itself. There is an Already have an account? Sign in link at the top, and the same link once more as the last line of the create-account card, below the form and the Google button, so arriving from the homepage as an existing customer is not a dead end. If you press Continue with Google on the sign-in screen and there is no Dante account for that Google address yet, we create one for you on the spot, under the Terms and Privacy Policy linked beneath the button, rather than stopping you with an error, and you carry straight on to the Knowledge step. If Google turns down the sign-up after you pick your account, or our server does, a message appears on the create-account screen instead of nothing happening. If you close Google's window, your browser blocks it, or Google does not answer, a line appears under the button after a few seconds: No answer from Google. If a pop-up was blocked or closed, allow pop-ups and try again, or continue with your email instead. If the Google button cannot load at all, the line says Google sign-in is not available right now. Please continue with your email instead. Either way, press Continue with Google again, or use Continue with email. On a phone-width screen, choosing Continue with email moves the Continue with Google button below the name, email and password fields, so the fields you are about to fill in are not pushed down past the fold; at desktop widths the layout is unchanged. The create-account card shows the Terms and Privacy Policy just once, in the line By continuing you agree to our Terms and Privacy Policy. If a name, email or password needs fixing when you press Create account, its message appears in space already kept under the field, so the fields below it do not move. The same is true the other way round. If you go to create an account with an email address that already has one, we do not stop you with an error: we take you to the sign-in screen with that address already filled in, and tell you why. Signing in there never throws away what you typed. You stay on the same screen and the website address stays in the box. If you had already pressed Start training and your account's agent is still in setup, we carry straight on and start reading that address on it, so you land on the training screen with nothing to redo. If your account is already set up, we do not add the address anywhere on your behalf: a line above the box tells you it was kept, and names the agent your next click will add it to. That choice stays yours, because adding a new site to an agent already answering your customers changes what it says to them. If you were a Dante customer before the new app, you do not need a separate address: sign in here with the email and password you have always used and we hand you straight over to your account. If a sign-in attempt is turned down, the card reminds you of this with the line Used Dante before? Sign in here with the same email and password. And if you decide not to sign in at all, the Back to the homepage link at the top of the card is the way out: the sign-in and create-account screens deliberately have no close button.
If your internet connection drops while you create your account, the button reads Waiting for your connection… and nothing is sent yet. If the connection comes back within 15 seconds, your sign-up goes through on its own. If it does not, we stop waiting, tell you nothing was sent, and the button changes to Try again, so you are never stuck on a disabled button. Reconnect, then press it. Nothing is sent later behind your back.
If you create your account with email, we send a code to confirm your address. You can leave the page to open your email app: if you come back to the same browser within 30 minutes, even after a reload, you land straight back on the code screen for that address rather than an empty sign-up form. We keep only your email address for this, never your password or the code. If it does not arrive, press Resend code. If a resend fails, a message says so and the button stays where it is, so you can try again without reloading. Once your address is confirmed, or after you continue with Google, Dante finishes setting up your account. If that takes longer than about 20 seconds, a Go to your dashboard button appears, so you are never left waiting on a spinner with no way forward.
Once your account is ready we send a short welcome email. In your inbox it shows a one-line preview of what the email is about, so you can tell at a glance it is the next step for your agent. The buttons and links in our emails about your account take you straight to the right screen in Dante, and each one carries a short label naming the email it came from, so we can tell which emails actually help people get set up and improve the ones that do not.
Already signed in? If you open the create-account page while you are signed in, it tells you which account you are signed in as and gives you two choices: Continue to dashboard, or Use another account, which signs you out so you can create a different one. It never starts a second account by accident.
2. Train it on your content
Paste your website address or upload a file on the Knowledge step. The first crawl is deliberately bounded: Dante reads a capped set of your site's core pages rather than the whole site, extracts the text, chunks it, and embeds it into your agent's private knowledge index. Only public pages are read, and your site is never changed. The exact page cap depends on how your agent was created and can change as we tune the first run, so treat the Knowledge screen as the source of truth: it lists every page that was actually read. If setup stopped because it reached the cap, a note under the list of pages read says so: it gives the number of pages setup reads and tells you to add more any time from Knowledge, so you know the rest of your site was not skipped by mistake and where to add it. The Knowledge step shows progress while it runs, so you can watch pages land, and the Sandbox opens as soon as your agent is ready. You can add more pages any time, up to your plan's memory limit.
Setup asks for one knowledge source before it hands you the dashboard, so your AI agent has something real to answer from the first time you talk to it. That is why there is no way to step past the Knowledge screen empty. A website is only one of three equal ways to meet it: the Website, Paste text and Upload file tabs sit side by side above the box, and any one of them is enough. If you have no site at all, Paste text takes your FAQs or a few paragraphs about what you do.
While you are in setup the top bar is deliberately pared back: the search box and its ⌘K shortcut are hidden, along with the Profile and Billing shortcuts in the account menu. Those all jump to parts of the product that do not exist for your agent yet, so during setup they were only ever a way to lose your place. They come back the moment setup is finished. For an extra AI agent that is still learning its first knowledge, search is only held back on that agent's own pages; the dashboard and the rest of the product keep it. This starts on the very first screen, before you have created an account, so a visitor trying Dante out is not dropped back onto the marketing site halfway through pasting their address. Signing out, contacting us and reporting an issue stay available throughout, so you are never held anywhere you do not want to be.
The top bar also has a Changelog button: the clock icon (on a phone it is in the account menu). It lists the product updates that have shipped, newest first, going back to April 2026.
Every label in the left sidebar shows in full, with nothing cut off: a long name wraps onto a second line instead. Where a row carries a mark on its right, such as the count of chats waiting for you or, where shown, a lock on a page your plan does not include yet, those marks line up in one column at the right edge of the sidebar.
If a crawl comes back empty, fails, or the site blocks our reader, the screen tells you what happened and hands the next move back to you: try another address, add one specific page, or switch to Paste text or Upload file on the same screen. You are never left on a finished screen with nothing to click. That includes retrying: if a second read fails too, the same screen comes back and says so, and once a read has failed a quiet Carry on and add content later link stays on the screen, even while a retry is running, so you can move into the product and add your content from there whenever you like. Each of those routes gets real content into your agent, which is the point: an agent that has read nothing of yours will still answer, fluently, from general knowledge, and that answer tells you nothing about how it will handle your customers. Getting your content in is worth the extra minute.
Sometimes the website box is already filled in when you arrive. That happens when we can work your site out from something you already gave us, either the address you entered before you signed up, or the domain in your email address (we never guess from a personal address like Gmail or Outlook). When that happens you see a short Is this your website? panel above the box that names the site and says where we got it from. Choose Yes, use this site to start training on it, or Use a different site to clear the box and type your own. Nothing is added to your agent until you choose, and you can always edit the address yourself. One arrival never fills it in: if you got here from the Build your agent free link under an AI agent on somebody else's website, the box starts empty on purpose, because the site you were just reading is theirs and not yours.
Uploading a file or pasting text is much faster than a website, usually seconds rather than minutes, because there is nothing to crawl. The screen names the file or the text you gave us rather than talking about pages, and the chat opens as soon as it has been read.
Reading your site usually takes a few minutes, and you do not have to wait it out. The screen names each page as it lands, so you can see what has been learned rather than watch a bar, and the chat opens as soon as the first page is in. Ask away while the rest arrives: the answers keep improving as more pages land. You can also give your AI brain a name while you wait: the screen stays put as new pages land, so the name you saved and the pages already listed stay where they are. Pressing Save on the name keeps the header and the top of the card in view, and the button that takes you on sits at the end of the card rather than floating over the page. The suggested questions above the message box are generated from your own content, so they are things your agent can actually answer. When you reach the question step, the question box is ready the moment the screen opens; only the suggested questions fade in after it. While Dante is still looking for pages, the screen keeps moving: a line under the title says what it is doing, or what usually comes next, such as connecting to your site or checking its sitemap, a clock shows how long it has been reading, and the page being read is marked reading now. If no page of your website has turned up after a short while, the card offers other ways in: Describe your business or Upload a document while the read carries on, or Start over with a different address. The rewards button is hidden while you set up, so it never covers the reading status at the bottom of the card. Each heading on the setup screens fades in once rather than being typed out letter by letter, and if your device is set to reduce motion it simply appears.
When the answer comes back, the screen names the source it actually came from: your site's address when the agent read your website, the document you uploaded, or the text you typed in. If you gave it more than one kind of source it says "what you gave it" rather than picking one of them, and if it has read nothing yet it names no source at all. That line is built from the same record as the list of what was read, so the two cannot disagree. If your agent replies that it does not have that information, the screen does not say it answered from your site: it says which pages it checked and that the answer was not there. That reply does not count as the answer that lets you carry on, so try another question. When the answer lists the pages it learnt from, each page name appears once.
How ready it is to go live. Under a real answer you see one line such as Ready to go live: four of five checks. Each check is something this step has already seen: that your website was read, that your own question got a real answer, and one check for each likely customer question Dante tried on your agent behind the scenes before suggesting it to you. Any check that did not pass is listed under the line as a next step, for example Teach it how to answer “Do you deliver on weekends?”. The line is only a guide. It never stops you publishing, and you can go live whatever it says.
3. See it work in the Sandbox
Open the Sandbox tab. Name your AI agent, then send a few test questions in the chat on the right and watch it answer from your own content. Once you have tested it, the Behavior panel on the left unlocks and you can pick a vibe (Friendly, Formal, Casual, Funny, and more) or write your own instructions. You'll watch the agent answer using your own content, so you can check its tone and accuracy before anyone else sees it.
4. Publish to your site
Setup finishes here. Pressing Let's go on the last setup screen takes you straight to the Publish page rather than to the dashboard, so the copy button and the snippet are the first thing in front of you, and Customize appearance and settings sits directly under them, already set to the Appearance tab, so your colours and branding are one tap away whenever you want them. The same screen lists other places the same AI brain can work, such as AI voice agent and AI in your inbox. On that screen, voice means talking to your agent out loud in the browser, and inbox replies are drafted for you to review before anything is sent. It does not offer a phone number or replies sent on their own.
On the Publish page, copy the embed snippet and paste it onto your site. The widget appears as a launcher in the bottom right corner.
Light and dark
Dante runs in light mode, so setup and the product look the same and nothing changes under you as you finish. Dark mode is on the way; you'll find it under Account, then Preferences, then Appearance when it arrives.
What happens after launch
Every visitor conversation lands in Chat history, under the Conversations tab. Your Free tier includes 100 credits a month, topped back up on the first of each month, so you can validate the agent answers the way you want. When traffic grows past that, the in-app upgrade prompt surfaces the cheapest unblock first (most users start with Starter at $40 a month for 2,000 credits, advanced styling unlocks at Advanced for $120, full white-label at Pro for $400).
On your dashboard, the big panel leads with what your agent has read, for example example.com is read · 1 page. A website that could not be read appears underneath it as its own row with the reason, a Fix button and a Remove button. The welcome card that offers a quick tour only appears on a workspace that has not been trained yet. If your workspace has no agent yet, that card offers Create your AI brain in place of the tour, because the tour walks through an agent: it opens the dialog to name your first agent. Get started closes the card either way. If you press Later on the tour, it stays closed for 24 hours, even after you close the tab, then picks up at the stop where you left it. The tour never opens by itself over a settings page or your account page, and it does not pull your cursor away from a field or button you are using. In a source's step details, ids and empty fields are left out.
Stuck? We're here.
Every plan, Free included, comes with human support. If your agent isn't behaving the way you expect after these four steps, send a note and someone on the team replies promptly.