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.

Creating your agent

Build your first agent through the onboarding wizard, then add more from your dashboard whenever you need them.

Open your dashboard in the app →

Your first agent

Your first agent is built through the onboarding wizard the first time you sign up. The wizard starts with your knowledge, so you can see the agent answering from your own content within a few minutes. This full step-by-step walkthrough runs only for your first agent. Every agent you add later still trains on your knowledge, but skips straight to the normal Sandbox once training finishes, see Adding more agents below.

1. Add knowledge first. On the Knowledge step, paste a website URL, upload a file, or paste your text in directly. You can start this before you finish creating your account. 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. If your site cannot be read, the Upload file and Paste text tabs on the same step meet the requirement just as well. If a read finds nothing and you type a phrase rather than a web address into the website field it offers, the message explains why and Type it out opens with what you typed, so you can save it as text. No website yet? If you type a few words about your business on the homepage instead of a web address, or press the button with the box empty, you are taken straight to creating your account. Afterwards your words are waiting in the text box where you describe your business, so you can add to them and save. A description of at least 50 characters gives your agent enough to answer from.
2. Create your account. Signing up spins up your first agent automatically and starts training it on the source you added. Training does not always succeed: some sites block automated readers, and if yours does, the Knowledge page says so and your agent has nothing to answer from until you add another source. If you are signed in but have no workspace yet, for example because you were removed from the only one you belonged to, setup makes you a new workspace named after you before it creates your first agent, so you are not stuck on We could not get started. If that still fails, the same screen keeps its Try again button.
3. Name your agent. You can name it as soon as reading starts, without waiting for training to finish. This is what visitors see in the chat header, so the name you type is the name that goes live with the widget. We suggest one based on your content, and typing your own always replaces it. You can change it any time. A green Saved. appears under the name only once you have saved it yourself and the save went through. If you type a different name, it goes away until you save that one too.
4. Test it in the Sandbox. Send a message, adjust the behavior, then retest with your real content. During setup, the last step asks you to put one question to your agent. If it cannot reply, the step says why. If your free messages have run out, or your agent has reached this month's message limit, asking again will not work, so the step offers See plans instead. For anything temporary, wait a moment and ask again. If finishing setup fails because your account's billing needs attention, the step links you to Update billing.
5. Deploy. Pick your colors, then embed the widget on your site or share a link.

Starting from a template

The templates gallery has a ready-made starting point for 20 kinds of business, from dentists and law firms to Shopify stores and property managers. Press Use this template on a card and you go through the same signup and Knowledge step as above. When your first agent is created, it starts with that template's welcome message, the three example questions from the card as starter questions in the widget, and a custom Behavior preset named after the template. The template only shapes how the agent talks: it still answers from the knowledge you add, not from the template.

Everything the template sets is yours to change in the Sandbox afterwards. A template is only applied to a brand new first agent. An agent that has already finished setup or gone live is never changed by opening a template link.

Leaving the wizard before you finish

Every step of the wizard has a Skip for now link at the top of the page. It takes you to your agent's Sandbox, where you can talk to it straight away.

Nothing you have already done is lost. Whatever pages have been read stay read, the name you gave your agent stays, and your agent is usable. From the Sandbox you can reach every setting the wizard would have walked you through, using your agent's own pages: Knowledge for what it reads, and Sandbox for testing and shaping how it answers. When you are ready to put the agent on your site, your dashboard links you straight to the publish screen.

This matters most if a step will not complete. If a step fails to save, use Skip for now rather than trying it again in a loop, and carry on from the Sandbox.

Right after your agent sends its first reply, pressing Skip for now opens one extra screen first: your embed snippet, a Copy snippet button, and an I've added it, verify check that confirms the moment your agent goes live on your site. Closing that screen for any reason, including I'll do this later, still takes you on to the Sandbox exactly as Skip for now always has, it is a single extra stop, not a new requirement.

Some accounts also see one extra screen when they press Skip for now on the Knowledge step. If you have not added anything yet, it offers the same choices as the step itself: Type it out or Use your website. If your site is already being read, it offers Ask it a question, which opens your agent's Sandbox. Pressing Skip for now on that screen carries on exactly as Skip for now always has.

Add a knowledge source first, then skip. Skipping is for after your agent has something to read. Once one source has been read, your choice is saved to your account, so it holds when you reload the page and when you sign in on another device. Skip before you have added anything and setup is still waiting: your agent has nothing to answer from yet, so the app takes you back to it the next time you move around.

You can come back, but nothing will remind you. The guided walkthrough still works after you skip it, and going back to it picks up where you left off. What is missing is a way to find it: nothing in the app links back, so you would have to know the address. If you want to be led through the remaining steps, it is simpler to finish them before you skip. Everything the walkthrough would have done can still be done from the pages above, just without the guide.

Adding more agents

Once you have an agent, you create more from the dashboard. Additional agents skip the guided step-by-step tutorial that your first agent ran. They open on their Website page, where you add what they should learn, with less setup since you already know the product.

1. Click New AI agent. The button sits at the top right of the Your AI agents list. You can also pick Add new AI agent from the agent switcher at the top of any agent page.
2. Name it. A dialog titled Name your new AI agent opens. Enter a name, then click Create. Free and Starter include one AI agent, Advanced includes 3, and Pro includes 10. If you are at your limit, the button does not open the naming dialog. On a paid plan it opens the Extra AI agent add-on ($20 / month each); on Free it opens the plans page, because add-ons need a paid plan. Either way the screen tells you it is the agent limit you have hit. Add what you need first, then the naming dialog appears.
3. Train it on your knowledge. The new agent opens on its Website page and starts with no knowledge, so paste a website URL there, or open Files to upload one, to begin training.
4. Land in the normal Sandbox. Once training finishes, the agent takes you straight to the normal Sandbox view, with no guided naming or testing walkthrough, where you can shape its behavior and deploy when you are ready.

Starting a new agent from a template. The Name your new AI agent dialog has a Start from list above the name box. Blank is picked by default and creates an empty agent, exactly as before. The other choices are Customer support, Lead qualifier, Booking assistant, FAQ and knowledge and Sales assistant. Picking one fills in a suggested name, unless you have already typed your own. Once you click Create, the new agent also gets that template's Behavior prompt, welcome message and starter questions. The template is only a head start: you can change all of it in the Sandbox afterwards, and the agent still needs your knowledge before it can answer.

Going to your AI agents when one is still empty. If you open the AI agents address (/agents) while your newest agent has not been trained yet, you land on your newest trained agent instead of the empty one, so you see an agent that already answers. To carry on with the new one, pick it from the agent switcher at the top of the page. Its training page says Start training your AI agent, not "your first", because your workspace already has a trained agent.

Building an agent by chat (being tested on some workspaces). Where we are testing it, New AI agent first asks how you want to set the agent up. Build by chat (recommended) creates the agent and opens a chat that walks you through its knowledge, a test question, its personality, its look and going live, with a preview of the widget beside it. Set up with forms opens the Name your new AI agent dialog and carries on exactly as the steps above. You can ask the chat to rename the agent at any step, and the new name shows everywhere at once. If you leave a chat build part-way, its row in the agents list says Continue building by chat and takes you back to it. Until you finish it, that agent's dashboard shows only the line This chatbot is still being built by chat with the same link: it does not say the agent is ready to put on your site, does not tick Add knowledge while the agent holds no characters of knowledge, and does not warn about a placeholder website such as example.com. On the last step, Open your agent's dashboard opens the agent you just built, even if you already had other agents: its page for trying it sits in the middle, with the usual menu on the left and the Backstage panel open on the right in the same chat as the build, so you carry on where you left off. You can close the panel and open it again from any of the agent's pages. On a computer the panel is always 400 pixels wide, and the bar at the top of the page is slimmer, so the agent has more room. On a phone the panel starts closed; tap Manage by chat at the top to open it in the same chat. It rises from the bottom of the screen and stops just under the top bar, so the agent's name and Manage by chat stay in view. It works before or after you tap Looks good, go live, which stays on the same step with your embed code and install steps. If you open the dashboard before going live, the dashboard says Continue building by chat to take you back to the build. If the builder has to try a step a second time, it does so on its own, and you only see a message saying a step did not work when its last try fails. If your workspace does not show the choice, nothing changes for you.

Starting a chat build from a file. The first step of a chat build asks for your website or a few words about your business. If you have neither to hand, press the paperclip in the message box, or drop a file on the box, and add a file about your business, such as a menu, price list or brochure. A file alone is enough: as soon as it is added, the builder starts writing your AI agent from it, without asking for a website or a description first. A file alone is enough: as soon as it is added, the builder starts writing your chatbot from it, without asking for a website or a description first. The paperclip takes PDF, Word, Excel, PowerPoint (.pptx), text, Markdown and CSV files, and PNG or JPG images; a file of any other type is named under the box and is not added. Each file becomes part of the agent's knowledge, the same as an upload on its Knowledge page, and starts training straight away. The paperclip stays in the box on the later steps too (asking a test question, personality and look), so you can add more files at any point in the build.

Trying your agent while it is still reading. In a chat build, the preview beside the chat opens as soon as the builder has made your agent, without waiting for your pages to finish reading, and you can type into it on every step after that, including Personality, Style and Go live. Until the first pages are read, the builder says it is reading your site and that you can already ask your AI agent a question. If reading stops before any page is in (the site could not be read, or the read went quiet for too long), the note says so and the preview stays open; send another address or add a file in the chat to teach it. When lead capture is on, the preview also shows the form your visitors will get, with your fields and intro message, so you can see it before going live; in the preview chat itself the form appears when your agent would show it on your site. Until reading is done, a note above the preview says Not read yet: you can try it now, and its answers get better once reading finishes. If you ask a question before anything has been read, the agent already answers from its instructions and anything you typed about your business, and says plainly when it does not know something yet; only an agent with nothing to go on at all says it is still reading what you gave it and asks you to try again in a moment. While your website is read, a box above the message box shows how many pages have been read out of the pages found, a short bar, the titles of the last pages read and a line naming one of those pages at a time. Once the first pages are in, it says you can try the agent in the preview while the rest is still read, and the note above the preview says how many pages it answers from so far. If the site blocks us or nothing could be read, the box says what happened and suggests pasting a few lines about your business, adding a file with the paperclip, or sending another address. If reading stalls, the box is hidden rather than left saying it is still reading. Everything you tell the builder about your business before it writes the first draft is saved together as the agent's About the business text, up to 1,000 characters; questions and short replies such as "ok" or "thanks" are left out. Each builder reply puts its answer first, with the steps Backstage took shown in grey underneath. When the builder or Backstage makes changes in reply to one of your messages, the changes are listed under the reply with one Change button: tapping it undoes every change in that message it can, newest first, and tells you the reason for any change it could not undo. You can undo a change for 30 days. Every reply that makes a change ends with a sentence or two on what changed and what to try next; the builder and Backstage do not describe each setting they read on the way.

Teaching it the right answer. When your agent says it does not know something, or you tell it an answer is wrong, a box opens under the question for the answer it should give. The box starts empty with a short prompt to write the answer your customers should get in your own words; it never fills in a sample answer from another business. When the builder already knows which common question your pages do not cover, such as opening hours, the prompt shows an example answer to that question. After you save, ask the question again to see the new answer.

Removing a source, and the draft catching up with your website. Under the message box of a chat build, each website, file or text your agent learns from is listed with a remove button, and you can also ask the builder to remove one ("remove example.com, that was a mistake"): it shows a Remove button for it. After you remove a source it disappears with Removed example.com. Undo for a few seconds; then it is deleted, your agent stops answering from it, and the builder updates the instructions without what came only from that source. This happens even if you leave the page or close the tab during those seconds. While the builder is making that update, a message you send is not sent: the chat says your agent is being updated after your source change, and you send it again a moment later. If your website was still being read when the builder wrote the first draft, the builder updates the draft once the site has finished reading, from what the pages really say, and tells you in the chat. It only changes words it wrote itself: anything you typed or edited in the instructions is left as it is. Each website is used for this update once.

Adding your website later in a chat build. If you started the build with a description, you can still add your website on any later step (asking a test question, personality, look or going live). Paste the address on its own in a message, for example https://yourbusiness.com, and the builder adds it to your agent's knowledge and starts reading it straight away, the same as on the first step, with an Undo if you change your mind. Pasting your website does not finish the step you are on. If the address is part of a longer sentence, the builder asks you to paste just the address.

Going live needs some knowledge first. Your chatbot only goes live once it has something to answer from: a website (one still being read counts), a file, a Q&A or a connected source, or a description of your business of at least 50 characters. A shorter description, such as "We sell shoes", is not enough, and a website or file that could not be read does not count. If you press go live without any of these, nothing changes and the builder asks you to paste your website address, add a file, or tell it about your business in at least 50 characters. On the go live step you can describe your business in the chat, and the builder saves your words as knowledge titled About the business.

Skipped steps and change cards in a chat build. The step bar at the top of a chat build shows a step you skipped with its number, a dashed outline and the word Skipped, never the done tick, and it still reads Skipped after you move on. When the builder offers a change, its card uses the same names as Settings: User message colour and Coloured header. A colour reads Solid #hex and a gradient reads Gradient #hex to #hex; if one colour would replace a saved gradient, the card says so before you apply it. To keep a gradient while changing one end, give the builder both colours. Knowledge limits on a card are written the way the Knowledge page writes them, for example 94,813 of 200M characters.

Changes Backstage could not make. When a change card is decided and one of its changes was not made, for example because a teammate had a draft open, the card says which change and why in one line, without opening it. Press Try again under that line and Backstage suggests the change again in a new message, so you can apply it once the reason has gone. While a card waits for you, each chat window preview on it shows all of the card's changes together: a new welcome message is shown under the new name the same card gives your AI agent. When you ask Backstage why your AI agent gave an answer, it reads that chat and checks which of your knowledge was trained at the time of the reply before it explains.

A refreshed look (being tested on some workspaces). Where we are testing our new visual style, the dashboard and the chat build use a light lavender page with white cards and one purple accent, and page titles in a serif face. On the chat build, the top of the page shows the Dante logo, the steps of the build and your profile picture or initials; the chat sits on a dark panel headed Backstage with the step you are on: Backstage's answers sit straight on the panel with no bubble, your own messages show in a typewriter-style face with a thin outline, the main buttons are lavender and the send button is a white square. The widget preview beside it keeps your agent's own light look, as your visitors will see it. On the chat build, the top of the page shows the Dante logo, the steps of the build and your profile picture or initials; the chat sits in a card headed Backstage with the step you are on, and the widget preview sits in a card beside it. Only the look changes: every step, button and setting works exactly as before. In the new style, once the preview opens, two tabs above it show your agent two ways: Full page (the default) shows it as a full-page chat, and Pop-up on your site shows the chat window at the size visitors see on your site, in the bottom right corner over your page. The round chat button in the corner opens and closes the pop-up, and so does the X at the top of the chat window. Your test conversation carries over when you switch. In that style the sidebar also has a Backstage box on every page except Backstage itself: type what should change and press Enter or the arrow, and Backstage opens and sends it straight away as the first message of a new chat, with no second press. If your workspace does not have the new style, you see the pages as they were.

How a chat build starts and ends. A new chat build opens with Backstage on its own, in the middle of the page, with a short welcome that says what you are about to build together: an AI agent for your website. The welcome ends by asking for your website or a few words about your business, and asks only once; if you gave your website before signing up, Backstage instead says it is already reading it and offers Write my AI agent. As soon as you send your first message (or add a file), the preview of your agent slides in beside the chat. If you come back to a build you have already started, you see the chat and the preview side by side straight away. When you press Looks good, go live, the install steps open in their own window: your embed snippet with Copy the code, how to add it to your site, Email this to my developer and the install check. Continue to dashboard takes you to your agent's pages with the Backstage panel open on the right, in the same chat, so you carry on where you were. The chat keeps a short Your agent is live line with Show install steps to open the window again.

When your website's security layer blocks us. Some websites sit behind a security service (for example Cloudflare) that refuses our reader. When that happens during a chat build, the builder tells you in the chat straight away, once, and offers three next steps you can tap: Paste my key pages (paste the text of your main pages into the chat and the builder adds it to your agent's knowledge), Upload a file (opens the file picker for a PDF, Word or other document) and How do I allow your reader? (the builder gives you the name our reader uses and the addresses it reads from, so whoever manages your site can allow it). You do not have to wait: the preview stays open and your agent answers from what you told the builder while the site is blocked.

Entering a website on the home page while signed in. If you are signed in and your workspace already has an AI agent, pressing Build my free AI agent on the home page first asks Add example.com to your agent?, naming the site and your agent. Add to followed by your agent's name reads the site into that agent's knowledge, so it starts answering from it. Cancel closes the question and nothing is read. To build a separate AI agent for the site instead, use New AI agent on your dashboard. If your workspace has no agent yet, there is no question: Dante builds your first agent from the site as before.

Switching between agents

The AI agent menu at the top of every page lists your agents. On the dashboard, picking one switches which agent the dashboard is about, and the menu on the left follows it, so its links always open the agent you are looking at. All your AI agents, at the bottom of that menu, takes you to the list of every agent on the dashboard.

Choosing how your agent behaves

There is no persona to pick when you create an agent. You shape behavior afterwards in the Sandbox. The Behavior preset dropdown gives you starting points, General AI agent, Customer support, Sales assistant, Product expert, and Lead qualification, or you can write your own custom prompt. See Instructions for how to edit the prompt.

What each agent gets

Each agent is a self-contained workspace asset. It has its own widget key, its own knowledge index, its own conversation log, and its own customization. Deleting an agent removes everything tied to it.

How many agents you can have

Free and Starter include one agent. Advanced includes 3 and Pro includes 10.

Need more, for example one support agent and one sales agent on the same workspace? Add the Extra agent add-on at $20 a month each. Each extra agent bills separately and gets its own widget key, knowledge index, and inbox. Add-ons need a paid plan, so Free users upgrade first. See Add-ons.

Stuck? We're here.

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.

Contact our team →