Publish your agent

Save vs Deploy, the embed snippet, allowed domains, the one-time deploy credit bonus, and the paid switches on the Publish page.

Open your agent's Publish page →

The welcome message

On the Setup tab, the welcome message is the first thing visitors see when they open the chat. Write a short greeting in your agent's voice, for example a sentence or two that says who you are and invites the visitor to ask a question.

• You can use up to 1000 characters. A live counter under the box shows how much you've used, and the box stops you at the limit so a long greeting always saves cleanly.
• If a save is ever rejected, the message under the box tells you exactly what to fix, instead of a generic error. Your edits stay in place and Save changes stays available, so nothing is lost.

Suggested questions and launcher teasers

On the Setup tab you can add quick replies: short suggested questions a visitor can tap instead of typing. They appear in two places:

• Inside the chat panel. Every quick reply you add is shown as a tappable chip, so visitors always see the full list once the chat is open. The chips lay out as a compact wrapped row rather than a stacked list, so a longer set stays readable without pushing the conversation down the panel, and the whole group settles in together instead of arriving one at a time.
• Beside the closed launcher. When Launcher tooltips is on, the first two quick replies also float as teaser bubbles next to the launcher icon before the visitor opens the chat, with a small unread count on the launcher that matches the number of bubbles shown. Keeping the floating stack to two bubbles stays unobtrusive on your site, so put your two strongest questions first in the list.

Tapping a teaser bubble or a chip sends that question straight to your agent. Visitors can dismiss the floating bubbles; they reappear for returning visitors after a day.

Your quick replies stand on their own. They show whenever you have added at least one, whatever else is switched on, and the AI follow-up suggestions toggle does not affect them: that setting only controls the extra questions your agent writes for itself after it has answered. The list you author on the Setup tab is the one visitors see at the start of every conversation. This is true on every install, whether you use the floating launcher script or drop the agent into your page as an iframe. Remember to Deploy after editing the list, since that is what pushes the change to your live site.

Behaviour and model

The Model tab controls how your agent answers. Pick a personality preset from the dropdown, or type your own instructions straight into the Behavior box: its tone, what topics to cover, and what to avoid. The Behavior box is always editable here, so you can refine the wording whenever you like, and Save changes turns active as soon as you make an edit. Saving stores your behaviour and model edits as a draft (see Save changes vs Deploy below).

The same tab is where you choose the AI model that powers your agent's replies. See AI models for how the models differ and what each one costs.

Save changes vs Deploy

These are two different things. Save changes writes your edits to a draft inside Dante AI. Nothing on your website moves. Deploy is the only action that pushes your saved edits to the live agent on your site.

• Save changes. Stores your setup, model, appearance, and behaviour edits as a draft. Visitors on your site still see the previously deployed version.
• Deploy. Takes your saved draft and pushes it to your live widget. Deploy asks you to confirm before anything goes live, and visitors see the change on their next page load.

On a phone or tablet

On a wide screen the Publish page shows your settings on the left and the live widget preview on the right. There is not room for both on a narrow screen, so below roughly 1024px wide the page shows one at a time with a Publish / Preview switcher at the top. It opens on Publish, so Save changes and Deploy are on screen straight away, and one tap flips to the preview to check how it looks before you ship it.

The settings tabs (Setup, Appearance, Model, Install, Handover, Tools) scroll sideways on a narrow screen, so swipe across them to reach the later ones. Everything you can do on a desktop, including deploying, you can now do from a phone.

After you deploy

When a deploy succeeds, a success screen confirms it on the spot: Your AI agent is live. It also hands you the two most useful next steps so you don't have to hunt for them:

• The embed snippet. If your AI agent isn't on your website yet, copy the snippet straight from the success screen and paste it into your site's HTML.
• Test it live. Opens your deployed AI agent in a new tab so you can send it a message and see exactly what visitors get.

If your plan still shows the "Powered by Dante AI" badge, the success screen closes with one quiet View plans link so you can remove Dante AI branding, add a custom domain, and see who is converting. It is optional, never blocks the next steps above, and does not appear once you are on Pro or already have the matching add-on active.

The success screen never locks you in. Close it with the X, the Done button, the Escape key, or a click outside the panel, and it quietly dismisses itself after a few seconds if you leave it untouched. Your deploy is already live either way.

If you deploy before adding knowledge

You can deploy an AI agent that has no knowledge sources yet, and nothing stops you. What changes is what your visitors hear.

An AI agent with no knowledge has nothing of yours to answer from, so rather than guessing, it gives every visitor the same short, honest reply:

This is deliberate. The alternative is an AI agent on your website answering questions about your business from general knowledge, which is how a live agent ends up stating things about your products, pricing, or policies that you never said. We would rather it say nothing than put words in your mouth.

• It never mentions Dante AI. Your visitors see a reply in your voice, with no mention of training, sources, or anything behind the scenes.
• It clears itself. Add a website, upload files, or paste text under Knowledge, and your AI agent starts answering from your content on the very next message. No redeploy needed.
• It can come back. If a scheduled recrawl later fails and leaves your AI agent with nothing to answer from, you will see the same reply again. Your dashboard tells you when this happens, so you do not have to notice it yourself.

In your sandbox you get a different, more useful version of this message with a direct link to add knowledge, because that screen is for you rather than your visitors.

If your live AI agent stops answering

A website can change after you go live. A firewall rule gets tightened, a security layer starts challenging automated visitors, or the pages we were reading move. When that happens, a scheduled recrawl can come back empty and leave a live AI agent with nothing of yours to answer from, even though it was working perfectly the day before.

You do not have to catch that yourself. On your dashboard, an AI agent in this state stops showing the green Live label and shows Live, not answering instead, with a note directly on the card that tells you:

• Which source is the problem, by address, so you are not guessing which one broke.
• What actually happened, in plain language. We could not connect at all, your site's security layer served us a challenge page, we found no readable pages, and so on.
• What to do about it, specific to that cause. Where the fix is allowing our crawler through, the note lists the exact addresses to allow.

We deliberately do not retry the crawl for you. If your site is turning us away, repeatedly knocking on the door does not help, and a fresh crawl uses your plan's character allowance. You decide when to run it again.

The notice clears itself as soon as your AI agent can answer from something again. That can be the original source once it crawls successfully, or a completely different one: uploading a document under Knowledge is often the fastest way back if your site is hard to crawl.

The deploy bonus

The first time your agent goes live on your own website, your workspace earns a one-time bonus of 500 credits. A few things to know:

• It is granted once per workspace, not once per agent. Deploying a second agent does not grant it again.
• It lands when your embedded widget actually loads on a real page, not at the moment you click Deploy. Paste the snippet on your site and open the page to trigger it.
• Code playgrounds do not count. If the widget first loads from a sandbox or playground host (for example CodePen, JSFiddle, CodeSandbox, StackBlitz, Replit, Glitch, JSBin, or w3schools), or from a localhost preview, your agent still goes live but the 500-credit bonus is held back. Embed on your real domain to earn it.

Two embed types

Pick the embed type on the Install tab. The snippet you copy changes with your choice.

• Floating AI agent. A launcher button in the bottom right that opens a chat window. Best for support and sales widgets on a customer-facing site.
• Iframe embed. A drop-in iframe you place anywhere on the page, for a dedicated chat page, a help centre, or an inquiry form. It defaults to full width and 600px tall, and you can change those values in the tag.

Allowed domains

Before you copy the snippet, add the hostnames where you'll embed it. The widget refuses to load anywhere else, which stops someone copying your snippet and running up your usage. The field also suggests hostnames pulled from your training sources so you can lock it in one tap.

• Hostnames only. No https://, no path, no port.
• www and non-www are covered together. Add example.com and we automatically also allow www.example.com (and the other way round), so your widget loads whether visitors land on the bare domain or the www version. You'll see both lines appear and a quick confirmation when this happens.
• Other subdomains count as separate hosts. Add both example.com and app.example.com if you embed on both.
• Leave it blank to allow the widget on any site. Useful during local testing.

Copy and install the snippet

On the Install tab, click Copy embed code. What you copy depends on the embed type you picked:

• Floating AI agent. A single <script> tag that carries your data-agent-id and data-widget-key. Paste it once, just before the closing </body> tag, on every page where you want the launcher.
• Iframe embed. A single <iframe> tag pointing at your agent. Drop it wherever you want the chat to appear.

You need the snippet copied at least once before Deploy unlocks. If you click Deploy before ever copying it, Dante AI sends you to the Install tab and highlights the snippet so you can grab it first.

Your widget key

The Install tab shows a widget key that authenticates the embed. Keep it private.

Paid switches on the Publish page

Two switches sit behind a plan gate. Toggling a gated switch opens the add-on prompt; the add-on activates once the payment confirms and the switch turns on, or upgrade to Pro to get both bundled.

• Dante AI branding. On the Appearance tab. By default the widget shows a small "Powered by Dante AI" badge; turn this switch on to hide it for a fully white-labeled widget. Bundled on Pro. On Free, Starter, and Advanced it is a $50 a month add-on.
• Human handover. On the Handover tab. Lets visitors ask to reach a person, routed to your team inbox. Bundled on Pro. On Free, Starter, and Advanced it is a $50 a month add-on, per 500 conversations.

AI search visibility

On the Install tab you'll find an Allow AI search engines to find this agent toggle. When on, AI assistants and search tools can read a public summary of this agent.

• Off by default. Nothing about your agent is exposed to AI engines until you turn this on. It applies the moment you toggle it, no separate Deploy needed.
• Your agent must be live. The summary only becomes available once your agent has been deployed on a real site. Unpublished agents are never exposed.
• Public-safe only. We share just four things: your agent's public display name, a short description of what it does, a few generic capability lines (for example "Trained on website content"), and the link to chat with it. Your persona or system prompt, your settings, your training URLs, and any private configuration are never included.

Roll back a deploy

Under the Publish header you'll see when the agent was last deployed, shown in orange. Use the Roll back menu next to it to restore any of your last 10 deploys. Rolling back loads that version into the editor as a draft, so it does not go live until you Deploy it again.

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 →