UXMachine

npx add-mcp 'https://uxmachine.app/mcp'

Connect it in two minutes

Pick your tool. The first time, a UXMachine window opens so you can authorize it with your account; if you don't have one, you can create it there.

Lovable

1. In your Lovable project, open Connectors → Add MCP server and enter this address: https://uxmachine.app/mcp
2. In the UXMachine window, sign in (or create your account) and click "Authorize".
3. In the project chat, ask: "Measure this project's home page with UXMachine." Lovable already knows which address to measure.

Step-by-step guide to connect Lovable

Claude Code

1. In your terminal, run: claude mcp add --transport http uxmachine https://uxmachine.app/mcp
2. Inside Claude Code, type /mcp and choose uxmachine.
3. Click "Authenticate". The UXMachine window opens in your browser.
4. Click "Authorize". Go back to Claude Code and ask for a measurement.

Step-by-step guide to connect Claude Code

Cursor

1. In Cursor, open Customize → MCPs → New MCP Server and paste this into mcp.json: {"mcpServers":{"uxmachine":{"url":"https://uxmachine.app/mcp"}}}
2. Click "Authenticate". In the UXMachine window, click "Authorize" and then "Open Cursor".
3. In Cursor's chat, use Agent mode.
4. When Cursor asks for permission to use medir and estado, click "Run".

How it works

You ask

"Measure the home page with UXMachine." Your agent sends us the address.

We measure

We open the page in a real browser, at 1280×800 and 390×844, on up to 11 pages of the site. It takes a few minutes.

Your agent acts

It reads the report, tells you what to change and, after the change, measures again.

What we observe, on real sites

Reports on public pages. They imply no commercial relationship, endorsement or approval.

Attio

The “email” field is declared as text. Judging by its name, it calls for an email address.

See the report →

PostHog

We found 1 menu that opens and closes without saying which of the two states it is in.

See the report →

See all the cases →

Free to start

You get 10 credits when you sign up. Each measurement uses 1. Up to 10 measurements a day, one at a time. Measurements we cannot make use none.

What each observation contains

• What is observed and where: the page, the window width and the element, with the measurement it comes from.
• The observable state to bring it to.
• How to check that it got there: which measurement runs again and which value counts as reached.
• If we cannot state something, we say so and say why.

Measuring again

When your agent measures again and passes us the earlier measurement, each rule comes back with one of four states: "Still observed", "No longer observed", "New" or "Not comparable", with its reason. "No longer observed" means it did not appear in the new measurement, not that it is solved. The comparison is by rule and page, not by element. You can compare your local version with the published one.

What it measures and what it does not

We measure what the page shows when it loads, in two windows (1280×800 and 390×844) and on up to 11 pages of the site. Only checks with an exact state that can be measured again the same way come out through this channel; they are on this list: /en/agents/rules. We do not fill in forms, we do not enter password-protected areas and we do not follow what happens after a click. We do not give opinions on design, copy or brand. This does not replace an audit or testing with people.

robots.txt and other people's sites

The channel respects robots.txt: if a site does not allow us to visit it, we do not measure it and we tell you so. Only ask for measurements of sites that are yours or that you have permission to review.

Your local version

If you work on your computer with Claude Code or Cursor, your agent can measure your site before you publish it.

• Your agent asks for a tunnel with the tunel tool, giving the port of your local server, and runs the command it gets on your machine. It only needs Node.
• The tunnel gives a temporary address that leads only to that port. We measure that address from our infrastructure, just like a published site.
• The tunnel closes by itself after 30 minutes. It only lets through requests that carry the key from the command, so someone with just the address cannot open your local version.
• The measurement is marked as a measurement of your local version, because it may differ from what is published.
• Development servers add their own layers, such as the Next or Vite error overlays. To get closer to what is published, measure a production build served locally.
• You can compare the local measurement with the one of your published site: your agent passes the earlier measurement when it measures again.
• Each measurement through the tunnel uses 1 credit, like any other.

Try it on your next change.