Let your agent scan any live UI.

One command connects uiscanner to Claude Code, Cursor, or Codex. Every scan returns real tokens, section structure, and a build prompt without ever leaving the chat.

claude mcp add uiscanner -- npx -y uiscanner-mcp

Works withClaudeCursorCodexTerminal

1
Open your terminal
In any project, nothing to clone first.
2
Run one command
No keys to paste, the first scan opens your browser to sign in.
```
claude mcp add uiscanner -- npx -y uiscanner-mcp@latest
```
3
Ask Claude
Sign in once, then tokens and a build prompt land right in your chat.
```
 ui_teardown example.com for a landing
```

Using Claude Code, Codex, or a plain terminal? The CLI is one command.

An agent building from a screenshot is guessing.

It eyeballs a hex, rounds a radius, guesses a type scale. uiscanner hands over the real ones.

01 · tokens

Real tokens, not guesses

Exact colors, fonts, type scale, spacing, and radii, pulled from the live page, never eyeballed from a screenshot.

02 · structured

Built for agents

Every scan returns structured output your agent acts on immediately: tokens, section-by-section structure, and a Claude-tailored build prompt.

03 · in chat

Scan mid-build, in chat

Tear down any URL without leaving your editor. The teardown lands in the conversation, no tab-switching, no copy-paste.

The command set

One scanner, a few verbs. The same engine your agents call.

uiscanner--help

scan
<url>[--target][--json]
Scan a public URL: tokens, section structure, and a build prompt.
probe
<url>[--json]
Pre-flight a URL: scannable, and is a recent teardown cached? No scan spent.
get
<id|link>[--json]
Fetch a finished teardown by its id or share link.
retarget
<id|link><archetype>
Re-tailor the build prompt to a new archetype, no re-scan.
whoami
[--json]
Show your plan and scans left this window.
ui_teardownin chat
urltarget
The MCP tools agents call in-chat: scan, get, retarget, whoami. No terminal.

FAQ

Do I need an account or API key?

You sign in once, with nothing to paste. The first scan opens your browser to connect your uiscanner account, and you're set. The free plan includes 5 scans every 30 days; Indie and Studio raise that to 100 and 400. There's no key to find or copy: we mint one for you on your plan. (For CI or headless runs, set a uix_ key in UISCANNER_API_KEY to skip the browser step.)

What does a scan actually return?

Design tokens (colors, fonts, type scale, spacing, radii), the section-by-section structure, and a Claude-tailored build prompt. We extract the system, not the brand: it describes and tokenizes the page, then links back to the original. No copied assets, no re-hosted bytes, and the work you build from the tokens is your own.

Can I make my AI app match my own brand?

Yes, and it's one of the cleanest uses. Scan your own marketing site once with ui_teardown, then point your agent at that teardown so every screen it builds matches your tokens, type scale, and spacing. retarget_teardown re-aims the same scan at a dashboard, settings page, or marketing layout without re-scanning. It's the opposite of cloning a competitor: you keep your own product consistent with itself.

Which agents work?

Anything that speaks MCP: Claude Code, Claude Desktop, Cursor, Codex, and others. No agent at all? Use the CLI in any terminal.

Can I use it in scripts or CI?

Yes. npx -y uiscanner-mcp scan <url> --json pipes the raw tokens into any build step.

Where does my data go?

Scans run against the hosted uiscanner API, same SSRF guard and privacy posture as the web app. See the privacy policy.