We built the world's first embeddable web agent - preview is live, roast us
Hey PH 👋
We're launching Rover on Feb 25th, but the preview is live right now and we want your honest takes before we go big.
What it is: One script tag on your website → your users get an AI agent that takes real actions inside your UI. Clicks buttons, fills forms, runs checkout, guides onboarding. Through conversation.
<script src="https://rover.rtrvr.ai/embed.js"></script>
Why we built it: Amazon spent billions building Rufus (embedded AI agent → $12B incremental revenue, 3.5x conversion lift). Google just dropped WebMCP asking websites to expose APIs for Chrome's agent. But 99.9% of websites can't build either. We made it one line of code.
How it works: DOM-native — Rover reads your actual page structure and acts on it semantically. No screenshots, no vision loops, no knowledge base to maintain. We're #1 on WebBench (81.39%) for a reason.
What makes this different from chatbots: Chatbots say "here's a link to our docs." Rover says "on it" and actually does the thing. Think: user says "connect my Salesforce" → Rover opens Integrations, clicks through auth, imports fields, builds the workflow. 40 seconds. Inside your UI.
Try it live → https://rover.rtrvr.ai
We're two ex-Google engineers who bootstrapped rtrvr from a hackathon project to 21,000+ users and ran ~1.5M+ workflows. Our mission is to build the Web agent infra that's highly accurate in its web data, extremely fast and cost effective. Would love feedback, questions, or your best roasts.
Subscribe for launch day → https://www.producthunt.com/products/rtrvr-ai?launch=rover-by-rtrvr-ai
Replies
this is genuinely impressive. the dom-native approach makes so much sense - i've seen too many "AI agents" that are basically just screenshot parsers with vision models, which breaks constantly.
few questions:
- how do you handle dynamic SPAs where the DOM changes constantly? (thinking React/Vue apps with heavy client-side routing)
- what's the latency like? 40 seconds for a full salesforce integration sounds fast, but curious about simpler actions
- security model? if this is running actions on behalf of users, how do you prevent malicious prompts from doing unintended things?
the timing is perfect with Google's WebMCP announcement. feels like we're at an inflection point where embedded agents will become table stakes for any serious web app.
good luck with the launch! will definitely be checking out the preview.
rtrvr.ai
@jonathan_song2Â
We actually havent had isssues the way we are DOM snapshotting with SPAs, the model sees the annotated DOM and has been able to figure what to act on or extract.
We have a benchmark listed here: https://www.rtrvr.ai/blog/web-bench-results showing that we were able to complete end to end tasks like booking an AirBnB on average less than a minute.
Security wise the website owner can setup blocklists for paths the agent cant interact with as well as additional prompt guidance. We also have an integrated platform where you can ground the agent with recording demonstrations and knowledge base of your domain.
Google's WebMCP protocol asks websites to expose their APIs so Chrome's agent can serve users directly. That hands Google the interaction layer. Rover is the counter-position; websites keep the agent and the user relationship on their own turf and not get further disintermediated by Google.