Building a Free Accessibility Scanner in One Week with an AI Agent Team: A Real-World Retrospective

Small web agencies often get that email from clients: "Is our website accessible?" There's no cheap answer to that question — a real answer requires a WCAG compliance audit that takes experts days to complete, and small teams can neither charge enough to make it worthwhile nor squeeze it into the gaps between projects. So the question gets shelved, or answered vaguely — which is somehow even worse.

Developer Sheree Preston's solution: use a team of AI agents to build a free, white-label accessibility scanning tool called Curbcut in roughly one week. She's refreshingly honest about what the tool can do, what it can't, and what it's actually like to "work with an AI team for a week."

A Question That Has No Cheap Answer

For small web agencies and freelancers, "Is our website accessible?" reveals a very real gap. Professional accessibility auditing is highly specialized work — it requires experts who live and breathe WCAG standards, and days of their time. That kind of expertise is hard to justify for a small shop: you can't charge enough for it, and you can't find the time to do it.

Sheree's product positioning is deliberately narrow and honest — not an audit, but a quick, plain-language "first look." You paste a URL, and about a minute later you get a scored, prioritized report: what's broken, which WCAG 2.2 rule it violates, who it affects, and what the suggested code fix looks like.

WCAG (Web Content Accessibility Guidelines) is an international standard developed by the W3C, with versions 2.1 and 2.2 being the most widely adopted. It divides accessibility requirements into three compliance levels: A (minimum), AA (typically the legal baseline), and AAA (highest). Each guideline maps to specific success criteria — for example, images must have alt text, color contrast must meet certain ratios, and all functionality must be operable via keyboard. In many countries, AA compliance has been codified into law — including the Americans with Disabilities Act (ADA) in the US and the European Accessibility Act in the EU — meaning non-compliant organizations can face litigation. That's why "Is our website accessible?" isn't optional from a business perspective; it's a potential legal liability that small agencies often lack the expertise to address.

The Core Design: White-Label + Honesty

Curbcut supports white-label output. Fill in the agency and client names, and the report header reads "Prepared for [Client] by [Agency]" — ready to hand directly to the client as your own deliverable. It answers that awkward question, opens the conversation, and points toward the remediation work that actually generates revenue — because fixing the issues is where agencies make their money.

The smart part of this positioning: the tool isn't competing with professional auditors. It exists as a lead-gen and conversation-starter. It acknowledges its own limits, leaves formal auditing to specialists, and takes on the role of "quick triage."

What It's Actually Like to Work With an AI Team for a Week

Sheree doesn't mystify the collaboration: she didn't write code, and she's not an accessibility expert. What she had was a team of AI agents — an "AI CTO" to translate her direction into a plan and break it into concrete tasks, and an "AI engineer" to implement, test, and verify. Her job was the parts that couldn't be delegated: describing the product in plain language, making decisions, and reviewing results the way a client would — reading reports, clicking through real sites, and writing blunt feedback.

Two incidents from the first week are telling. First, she accidentally pasted an email address into the URL field and got a confusing rejection message. She sent a screenshot; the error message was rewritten the same day. But digging into the cause revealed a bigger gap: standard tools can't scan password-protected staging sites — which is exactly where agency work happens. That same week, the team shipped support for "staging links with credentials in the URL," with a strict rule: credentials are only sent to the exact host you pasted, and never appear in reports, logs, or anywhere else.

Second, she tested a site whose firewall blocked automated visitors. The scanner correctly reported a failure, but the message was too vague to tell if it was a bug or a site-side issue. The prompt now explicitly states "the website blocked the scan" and explains why this is likely the cause when the page opens fine in a browser.

The loop was simple: she broke it or described it, the team fixed it, and nothing shipped without passing the green automated test suite and real-world checks. When she got the first all-green report, she didn't celebrate — she asked the tool what it wasn't checking. That question changed the product again.

The "AI CTO + AI engineer" architecture Sheree describes is a concrete example of the multi-agent collaboration model gaining traction in AI-assisted development. Unlike a single conversational AI, a multi-agent framework assigns different responsibilities to different roles — one translating ambiguous requirements into actionable technical plans, another handling code implementation and testing. This structure mirrors human team dynamics and reduces the context drift and task bleed that often plague single-model approaches. For non-technical founders, the key isn't understanding the code — it's being able to articulate clear product intent and apply rigorous acceptance criteria to results. That's exactly what Sheree emphasizes throughout. Worth noting: the 333 automated tests weren't written by hand — they accumulated incrementally through the agent team's iterations, which is itself one of the structural safeguards that keeps AI-assisted development from degrading in quality over time.

What Curbcut Can Do Today

The capability list is specific and measured:

• Paste a URL and go — no signup, no email, no credit card required; supports password-protected staging links.
• Fetches the target page and up to four linked pages on the same site, running a fixed set of 14 specific WCAG 2.2 checks against the HTML the server returns.
• Delivers results in about a minute: a scored, prioritized, plain-English report with suggested code fixes.
• If a check can't make a determination about a page, it's listed rather than guessed — displayed in an amber "What Curbcut Couldn't Check" panel below the score, so a green result actually means something.
• Reports are white-label and print-ready for direct client delivery.
• Privacy-first design: the server retains only a bare count of daily scans — no URLs, no visitor data, no cookies.
• Every change ships under an automated test suite (333 passing tests at time of writing) and has been verified against real sites including gov.uk.

What It Can't Do — and Says So Explicitly

This is what Sheree cares most about, because overclaiming is exactly how trust collapses.

• It reads the HTML the server sends. It can't see what the browser renders afterward via JavaScript — carousels, cookie banners, embedded widgets. A heavily JS-dependent page may score better than it deserves.
• Two checks — color contrast and focus indicators — require the site's stylesheet. When the stylesheet can't be read, they're skipped, and the amber panel says so plainly rather than guessing.
• It scans one site at a time, a few pages deep. It's a report generator, not a crawler.
• Most importantly: this is not a compliance audit. It doesn't certify sites or provide legal protection. It's the quick, prioritized first look that opens the door to real work — and if a client needs a formal audit afterward, that specialist is still worth every penny.

The gap between static HTML scanning and JavaScript-rendered content is the most fundamental technical limitation facing automated accessibility tools. Modern web pages rely heavily on client-side JavaScript to dynamically generate the DOM — single-page applications (SPAs) built on React, Vue, Angular, and similar frameworks present content to users that looks almost nothing like the raw HTML the server returned. For heavily frontend-framework-driven sites, a static HTML scanner can miss large classes of real accessibility issues: dynamically loaded modals, interactive components without focus management, missing ARIA attributes injected at runtime. More mature tools like axe-core and Lighthouse^Ad partially work around this by running in a real browser environment, but that introduces significantly higher operational costs and complexity. Curbcut's choice to use static scanning — and to openly declare this limitation — is a deliberate tradeoff between engineering cost and honest capability boundaries.

The Current State and What's Next

Sheree is precise about where feedback has come from: everything that has shaped the product so far comes from internal testing — her own scans, her own mistakes, and the team's automated checks and real-site verifications. No external users have weighed in yet, because there aren't any external users yet. That's exactly the point of this stage: real-world testing, feedback-driven iteration. No roadmap theater — the next feature will depend on what the first real users say is missing.

What makes this case worth paying attention to isn't just the tool itself — it's the "one person + AI agent team" collaboration model it demonstrates: a non-technical founder, through clear product description, decisive judgment, and rigorous result evaluation, shipped a usable product in one week with 333 tests and real-site validation. It shows both the productivity ceiling of AI-assisted development and honestly marks the boundaries of what the tool can do. That honesty, as it turns out, is the rarest quality of all in this category of product.