# I built an accessibility scanner with an AI team in a week — and it's free > Source: > Published: 2026-10-11 21:52:38+00:00 Small agencies keep getting asked "is our website accessible?" — and a specialist audit is out of reach at small-shop rates. So I built a free, white-label scanner with a team of AI agents in about a week. Here's how the work actually went, and what my tool honestly can't do. TAGS (dev.to allows four): accessibility, webdev, a11y, buildinpublic BODY: If you run a small web agency, or freelance for a living, you know this email. A client has read something, a procurement form asked, or a competitor got called out — and it lands in your inbox: *"Quick question — is our website accessible?"* There's no quick answer. The real one is a conformance audit — specialist work, days of it, by someone who does WCAG for a living. That's expertise a small shop can rarely justify: you can't sell it at your rates, and you can't credibly do it between builds. So the question sits unanswered, or gets answered vaguely, which is worse. I'm Sheree. I run Curbcut, and earlier this month I built it with a team of AI agents in about a week. The idea was deliberately small and honest: not an audit — a fast, plain-English first look. You paste a URL, and about a minute later you get a scored, prioritised report: what's broken, which WCAG 2.2 rule it breaks, who it hurts, and suggested code fixes. It's white-labelled too: fill in your agency and your client's names, and the report is headed *"Prepared for [client] by [agency]"* — you hand it over as your own work. It answers the question, starts the conversation, and points at the fix work — where the real money is for an agency anyway. No mysticism: I didn't write the code, and I'm not an accessibility specialist. What I have is a team of AI agents — an AI CTO who turns my direction into a plan and delegates concrete tasks, and an AI engineer who implements, tests and verifies. My job is the part that can't be delegated: describing the product in plain words, making the calls, and judging results the way a client would — reading reports, clicking through the live site, sending blunt feedback emails. Two episodes from that first week tell you how it goes. The first: I pasted an email address into the URL box by mistake and hit a confusing refusal. I sent a screenshot. The message was rebuilt the same day — and chasing why it happened surfaced a bigger gap: standard tools can't scan password-protected staging sites at all, which is exactly where agency work lives. That same week we shipped support for staging links that carry their sign-in in the URL, under strict rules: credentials are sent only to the exact host you paste, and never appear in the report, the logs, or anything else. My mistake became the feature I'm proudest of. The second: I tested a site whose firewall refused automated visitors. The scanner correctly reported failure — but the message was too vague to tell whether it was a bug or the site. The message now says plainly that the website refused the scan, and why that's the likely cause when the page opens fine in your browser. Even the failure path should be honest. That's the whole loop: I break it or describe it, the team fixes it, and nothing ships without a green automated test suite and a live check. When I got my first all-green report, I didn't celebrate — I asked what the tool had *not* checked. The answer changed the product again. The facts: It's free while I work out whether it deserves a price tag. **Try it here: [https://05910e406f67ad386ac82700fa809394.ctonew.app](https://05910e406f67ad386ac82700fa809394.ctonew.app)** This is the part I care about most, because overclaiming is how trust dies. The build took about a week. Whether it's useful is decided by people scanning real client sites. To be precise about where feedback has come from so far: everything that shaped the product came from internal testing — my own scans, my own mistakes, and the team's automated checks and real-site sweeps. No outside users have weighed in yet; there aren't any yet. That's exactly what this phase is for: real-world testing, feedback-driven. No roadmap theatre; the next feature will be whatever the first real users say is missing. Curbcut is free: paste a URL, get the report, no signup. Take a site you know well — a deliberately rough one is the most instructive — and see if you'd hand the report to a client. Then tell me what it got right and what it missed, in the comments here or through the contact link on the site. Blunt is fine. It's how this thing got built in the first place. — Sheree Preston, founder, Curbcut