Works on any site. Your users install nothing.

The AI that does it,
not just explains it.

Every other chat widget writes your user a set of instructions and leaves them to it. AgentWidget reads your live page, works out the steps, and does them — clicking, typing, navigating — then checks its own work before it says it's done.

One script tag. No content to train, no flows to configure.

Install

One line, wherever your product lives

Paste it and you're done. There's nothing to feed it — it reads the page the way your user sees it.

<!-- paste before </body> --> <script async src="https://cdn.agentwidget.io/v1.js" data-key="aw_pk_live_YOUR_KEY"></script>

Works in any HTML page, template, or server-rendered app.

// app/layout.tsx — inside <body> import Script from 'next/script' <Script src="https://cdn.agentwidget.io/v1.js" strategy="afterInteractive" data-key="aw_pk_live_YOUR_KEY" />

App Router or Pages Router. afterInteractive keeps it off your critical path.

// index.html in public/ — or useEffect if you prefer <script async src="https://cdn.agentwidget.io/v1.js" data-key="aw_pk_live_YOUR_KEY"></script> // Then anywhere in your app: window.AgentWidget.open('how do I export this?')

Vite, CRA, Remix — anything that ships an HTML shell.

// functions.php in your theme add_action('wp_footer', function() { ?> <script async src="https://cdn.agentwidget.io/v1.js" data-key="aw_pk_live_YOUR_KEY"></script> <?php });

Or drop the plain HTML snippet into any header/footer scripts plugin.

<!-- Online Store → Themes → Edit code → theme.liquid --> <!-- paste just before </body> --> <script async src="https://cdn.agentwidget.io/v1.js" data-key="aw_pk_live_YOUR_KEY"></script>

It can walk a shopper through checkout, filters, and returns.

// Settings → Custom Code → Add code to Body - end <script> window.AgentWidgetSettings = { key: 'aw_pk_live_YOUR_KEY' } </script> <script async src="https://cdn.agentwidget.io/v1.js"></script>

Wix strips some attributes, so pass config via the settings object.

<!-- Project Settings → Custom Code → Footer Code --> <script async src="https://cdn.agentwidget.io/v1.js" data-key="aw_pk_live_YOUR_KEY"></script>

Publish after saving — custom code only runs on published sites.

Live

This is a real product, not a video

A working dashboard with AgentWidget on it. Ask it to change a setting, find a customer, or explain a page — it will actually go and do it.

acme-saas.demo

How it works

Read the page. Do the thing. Check it worked.

1

It reads your live page

No training, no docs to upload, no flows to record. It sees exactly what your user sees, including whatever you shipped this morning.

2

It does the steps

Clicks, types, picks from dropdowns, moves between pages. Anything your user could do with a mouse and keyboard, it can do for them.

3

It verifies

After acting it looks again. If the page didn't change the way it expected, it tries another route instead of claiming success.

The difference

Telling isn't helping

A normal AI chat widget

  • "Go to Settings, then Billing, then click Manage Plan…"
  • Needs your help docs ingested and kept up to date
  • Goes stale the moment you ship a UI change
  • Deflects the ticket; the user still has the problem
  • Billed per resolution, uncapped

AgentWidget

  • "Done — I've switched you to the annual plan."
  • Reads the live DOM, so there's nothing to ingest
  • Ships when you ship — it sees the new UI immediately
  • Finishes the job, so activation and retention move
  • Flat monthly. No per-outcome meter

Safe by default

It asks before anything it can't undo

Confirms destructive steps

Delete, pay, send, publish, cancel — it highlights the button and waits for a yes. Twice over: the model is told to ask, and the widget blocks it regardless.

Guide mode

Not ready to let it act? Run it in guide mode and it spotlights the right button and explains, but never clicks. Locked server-side, so it can't be overridden.

Locked to your domain

Your key only works from domains you register. It never reads password or card fields, and page content can't issue it instructions.

Pricing

Flat monthly. No per-answer meter.

Starter

$19/mo

For a single product

  • 500 tasks a month
  • 1 domain
  • Guide + do modes
Get early access

Pro

$49/mo

For a growing SaaS

  • 2,500 tasks a month
  • 5 domains
  • Transcripts and analytics
  • Custom branding
Get early access

Scale

$199/mo

For a real user base

  • 15,000 tasks a month
  • Unlimited domains
  • Priority support
Get early access

For comparison: Intercom Fin charges $0.99 per resolution with no cap — about $1,485/month at 1,000 resolutions, and that's for answering. AgentWidget's Pro plan covers 2,500 tasks it actually completes, for $49. Fin's pricing.

Questions

The things people ask first

Do my users have to install anything?

No — and that's the whole point. Browser agents like Gemini in Chrome only help people who've already installed them. Almost none of your users will. AgentWidget is already there for every single one, because you installed it once.

Will it break when I redesign?

No. It reads the live page every time rather than following a recorded script, so a redesign is just a different page to read. There's nothing to re-record and no content to re-ingest.

Can it do something destructive by accident?

It stops and asks before anything hard to undo, and the widget enforces that independently of the model — so even a page trying to trick it can't get a delete through without your user clicking confirm. You can also run it in guide mode, where it can only point.

What about my users' data?

Password, card and OTP fields are never read. Your API key stays on our server and never reaches the browser. You can mark any element data-agent-ignore and it becomes invisible to the agent.

Does it work in a single-page app?

Yes. React, Vue, Next, Remix — it types through the same events your framework listens for, so controlled inputs update properly, and it re-reads the page after every step so it never acts on a stale render.

What counts as a task?

One thing your user asked for, however many steps it takes. "Invite three teammates and set their permissions" is one task, not eleven.

Put it on your product

Early access is going out in small batches so we can help each team get it right.