Use case · developers
Point it at a URL and it goes through the page the way a careful reviewer would — broken images, dead links, console errors, layouts that fall apart on a phone. What it finds arrives as pins, marked as its own, for you to keep or delete.
Have it swept first.
Ask it in the chat you already use. Your agent opens the page with the browser tool it already has; Handoff is where what it finds gets filed.
Broken images, dead links, console errors, text that overflows its container, layouts that collapse at narrow widths — the things that are obvious once seen and easy to never look at.
Each one carries a screenshot, a selector and the page it's on, and is filed as the agent's own, so it reads as a suggestion rather than a colleague's judgement.
Keep the ones that matter, delete the noise. The agent's opinion is a starting point for the review, not the verdict.
The value isn't that an agent sees more than you do — it's that it will look at the page for the tenth time this month with exactly the same attention as the first, which no person will.
Claude Code · via MCP: “Audited /pricing — 3 findings: hero image 404s, the FAQ toggle throws on click, plan cards overflow at 375px.”
Four pins on a demo product, all posted by an agent through create_comments — contrast, a form error in the wrong place, a header that wraps, and an empty state written in database language.
Each finding became its own review, carrying the element screenshot the agent captured.
An AI page audit is a pass over a live page by an agent, looking for the things that are obvious once someone notices them: assets that stopped loading, links that lead nowhere, errors in the console, text that overflows, layouts that break at a width nobody tested.
None of this is hard to spot. That's precisely why it goes unspotted — checking is dull, gets skipped under deadline, and the page most likely to be broken is the one nobody has opened in a month. Automated checks exist, but they need writing and maintaining, and they only ever cover the failures somebody thought of in advance.
An agent sits between the two. Given a way to open the page — a browser tool most coding agents already have — it reads the page as rendered rather than as source, so it sees what a person would see, and it will do it again next week without needing to be reminded. What it produces is a list of suspicions, filed as pins with a screenshot and a selector and marked as its own — because the judgement of which ones matter stays with a person.
Sometimes — it files suspicions, not verdicts. Findings are marked as agent-created and sit in the same list as human notes, so you keep the ones that matter and delete the rest.
No. It catches what's visible on the rendered page, which is a different set of failures from what a test suite catches, and it needs no maintenance to keep doing it.
It audits what it can reach. A public page or a preview build works directly; a page behind a session is a job for a person with the extension.
Whichever MCP client you already use — Claude Code, Cursor, Codex. The audit is a request in your normal chat, not a separate product.
On the Task Board they carry an AI-agent label, so a machine's guess never looks like a colleague's judgement.
Your agent needs a way to open pages — a browser tool most already have. Handoff doesn't drive the browser; it's where the findings land, with a screenshot and a selector, next to your team's own notes.