A/B testing, run by your AI agent.A/B testing, run by
ClaudeChatGPTCursorCodexyour agent

A website you can add one script to (or an agent that can, through your CMS or code) and an AI agent that speaks MCP: Claude, ChatGPT, Cursor or similar. Your agent connects to Duelpilot with one prompt.

The script is 14 KB from a global edge network and runs before the page paints. Only the elements under test wait, usually for a few milliseconds; the page itself is never hidden. In our tests it costs about one network round trip.

If visitors see it in a browser, almost certainly. Duelpilot works on the finished page, so the framework doesn't matter: React and Next.js, Vue and Nuxt, Svelte, Astro, Turbo, single-page apps and sites with strict security headers are covered by our browser tests in Chrome, Safari and Firefox, and WordPress, Shopify, Webflow or Wix pages are ordinary HTML to it. What it can't change in the browser: content inside other companies' widgets (embedded iframes) and anything your server decides, like prices or search results. That is what code tests are for.

It finds them itself: the Analyst reads your analytics and the page structure and puts ideas on the board with evidence; the Designer inspects the page and checks every selector before it submits. You can point too: open 'Pick on my site' from your board and click any element on your live site to turn it into an idea, or use Point & comment on a preview to say 'make this bigger'. The pick link only works for signed-in members, expires after 30 minutes and only creates ideas, so nobody can change your site with it.

Yes for what most tests change: text, colours, styling, images, hiding and reordering keep the original element, with everything it does. New blocks a variant inserts are plain HTML, so a new button is a link, never a new script. Before anything can go live, the site check clicks the buttons, menus, date pickers and form fields in and around the change on phone, tablet and desktop, without sending anything, and compares the variant with the original: if something stopped responding, the check fails and you see it before you approve.

Changes in how something works (a new flow, a different date picker, fewer form fields, pricing or search logic) belong in your code. On the Code plan your own agent builds each variant behind a Duelpilot flag in your repository and opens the pull requests with its own GitHub access; Duelpilot never needs access to your code. Duelpilot splits the traffic, checks every variant's preview, measures, and when a winner ships your agent opens the clean-up pull request.

Not unless you want it to. By default a person approves every test on the board. An owner can turn on autopilot with guardrails: which pages, which kinds of changes, how much traffic. Harmful tests pause themselves either way.

Duelpilot stores no visitor profiles and no personal data. It follows your consent setup (Google Consent Mode, TCF or your cookie banner) and has a cookieless mode. Whether it may run before consent is your call; your agent asks you.

Nothing extra from us. The agents run in your own Claude (or other agent) subscription; Duelpilot never resells AI.

GA4 is set up automatically, including results back into Duelpilot. Variant events also go to Google Tag Manager, Adobe, PostHog, Mixpanel, Amplitude, Segment, Hotjar, Clarity, Plausible, Fathom, Matomo and datafa.st automatically; anything else gets them through a one-line bridge.