How can my AI coding agent check my website before saying it's done?

Give it a tool that measures the page in a real browser, and ask it to report three things: what it checked, with what evidence, and what it could not check. UXMachine is one such tool: an MCP server your agent calls after changing what people see.

Why "done" needs a check

An agent can confirm that the code builds without having opened the page at a phone's width, or checked whether a new button has a name.
A report that only lists what was found makes a page that could not be checked look clean.

What to ask your agent

Measure the changed pages before saying the work is finished.
Say where each observation was seen and how to check it again.
Say what could not be checked, and why.
After a fix, measure again and compare with the previous measurement. "No longer observed" means it was not seen this time, not that it was fixed.

How UXMachine does it

Your agent connects to https://uxmachine.app/mcp (setup steps), calls medir with the address, then estado until the report is ready.
It opens up to 11 pages at 1280×800 and 390×844 and runs a fixed, public list of checks. No AI model takes part in the check.
It returns observations with evidence, and what it could not check, with the reason.
Given the earlier measurement, it reports each check as new, still observed, no longer observed or not comparable.
For a version running on your machine, the tunel tool lets the agent measure it before you publish.
A GitHub Action can also measure after your deployments (by default, at most once every 6 hours).

Limits

It checks the page as it loads: it does not fill in forms, enter password-protected areas or follow what happens after a click.
It does not measure speed or give a score, and it does not replace testing with people.
A measurement uses 1 credit only when it returns a report with content.

This page describes documented features, not results on your site. Checked on 20 October 2026.