Miromiro MCP Server

miromiro-mcp

MCP server for the MiroMiro design-extraction API. Give your AI agent the real design of any live website - brand colors, fonts, design tokens, SVGs, images, Lottie animations, and paste-ready component code - so it builds from actual values instead of guessing from a screenshot.

Works with Claude (web, desktop, Claude Code), ChatGPT, Cursor, Windsurf, Codex, and any MCP client.

Claude Code plugin

Install as a plugin to get the tools plus five skills that teach Claude when to reach for each one:

• elementor-section - any site's section as an Elementor template to import into WordPress
• shopify-section - any site's section as a Shopify theme section (.liquid)
• clone-section - a section as Tailwind, HTML+CSS, React or Vue
• match-brand - a site's colors, fonts and design tokens to build in its style
• extract-assets - images, SVG icons, videos, audio and Lottie files

claude plugin marketplace add SoraiaBarroso/miromiro-mcp
claude plugin install miromiro@miromiro

The plugin uses the remote server, so there is nothing to install and no key to paste - on first use it opens a one-click MiroMiro sign-in that links a free account (300 credits/month, no card).

Skills for OpenClaw and other agents

The same skills live in skills/. Each one uses the MiroMiro MCP tools when they are connected and otherwise calls the REST API with MIROMIRO_API_KEY (free key, 300 credits a month: https://miromiro.app/api/onboarding).

Quick start: remote server (nothing to install)

A hosted server lives at:

https://miromiro.app/mcp

Paste that URL into any client that accepts a remote MCP server. On first use it opens a one-click MiroMiro sign-in that links a free account (300 credits/month, no card).

• Claude (web or desktop): Settings → Connectors → Add custom connector → paste the URL. Or use the one-click install.
• ChatGPT: Settings → Security and login → enable Developer mode, then Settings → Plugins → New Plugin with the URL (authentication: OAuth).
• Cursor (remote): add { "mcpServers": { "miromiro": { "url": "https://miromiro.app/mcp" } } } to ~/.cursor/mcp.json.
• Claude Code: claude mcp add --transport http miromiro https://miromiro.app/mcp, then /mcp in a session to sign in.
• Codex: add [mcp_servers.miromiro] with url = "https://miromiro.app/mcp" to ~/.codex/config.toml, then run codex mcp login miromiro.
• Grok Bot: ask the bot "Add a custom MCP server called miromiro at https://miromiro.app/mcp", then press Authorize.
• Muse: Settings → Connectors → Add custom connector → paste the URL.
• Step-by-step pages for each agent: https://miromiro.app/api/agents/chatgpt (swap the last part for claude, claude-code, codex, cursor, grok-bot, muse or openai-dots).
• Raw API key instead of OAuth: append ?access_key=mm_live_... to the URL, or send it as an Authorization: Bearer header.

Local server (npm)

Prefer a local stdio process with the key in an env var? Get a free key at miromiro.app/api/dashboard, then:

Cursor (~/.cursor/mcp.json) / Windsurf (Settings → MCP):

{
  "mcpServers": {
    "miromiro": {
      "command": "npx",
      "args": ["-y", "miromiro-mcp"],
      "env": { "MIROMIRO_API_KEY": "mm_live_your_key" }
    }
  }
}

Claude Code (run in your terminal, not inside a claude session):

claude mcp add miromiro -s user \
  --env MIROMIRO_API_KEY=mm_live_your_key \
  -- npx -y miromiro-mcp

Tools

All tools are read-only and retry-safe. Identical calls within 24 hours are served from cache for free, and failed calls are never billed.

Asking for a part of the page by name

extract_code takes section - one of hero, header, footer, pricing, features, testimonials, faq, cta - so the agent never has to invent a CSS selector for a page it cannot see. Every result also lists the page's other sections with ready selectors for follow-up calls. A selector or section that matches nothing falls back to the hero with a note saying what the page does have, instead of a dead-end 404.

Every extract_code result carries previewUrl: a page showing exactly the returned code rendered, so you can compare it with the live site even if your agent adapts the code; downloadUrl: the code exactly as returned, as a file in the format you asked for (.jsx, .vue, .html, .elementor.json, .liquid); and pagePreviewUrl: every section of that page extracted so far, assembled in order. Agents are told to build on the returned code rather than regenerate it: they make it better to use where they can (clear names, reusable components with props, shorter code without redundant wrappers or duplicated classes, your theme tokens) as long as it renders the same, hand clean code over as it is, and never retype the real layout and styles. Inline SVGs over 2 KB (logo walls, illustrations) come back as permanent hosted .svg files behind an <img> that keeps their classes, size and color, which keeps the code small enough for the agent to use as-is (astro.build's hero: 73 KB to 14 KB). Pass inline_svg: true for fully self-contained code.

Prompts (slash commands)

The server exposes three MCP prompts, which clients surface before the first call - Claude Code as /mcp__miromiro__clone_section, /mcp__miromiro__brand_kit and /mcp__miromiro__design_tokens, VS Code as /miromiro.<name>, Cursor in its prompt menu:

Controlling output size (remote server)

Remote results are trimmed by default so a big page cannot flood the chat context, and every trimmed response says exactly what was dropped. Two optional arguments loosen it per call:

• response_format: 'concise' | 'detailed' - the one-word knob: concise keeps lists to 10 entries for a chat answer, detailed returns everything the tool has
• limit (up to 200) - raise the item cap: "give me 80 SVGs, not 20"
• full_output: true - complete long strings, like entire inline SVG markup

Generated code is never trimmed. Results stop at roughly 20k tokens (25k for extract_code) so they fit the tool-result cap of Claude Code, Claude Desktop and Cursor, and the truncation_note says how many tokens were cut and which argument brings them back.

The REST API and this local npm server always return everything untrimmed, so limit and full_output apply to the remote server only. Every tool on both servers accepts render: true (forces a real browser, +5 credits, on every plan; usually not needed because the server re-runs JavaScript-only or refused pages in a browser by itself - see Notes), plus two more shared arguments:

• fresh: true - skip the 24-hour cache and re-extract the live page (charged normally); use when the site just changed
• screenshot: true (requires render: true) - the result gains a screenshot field: a public URL of a 1440x900 capture of the page, live for 7 days, at no extra credits - show it to the user to verify an extraction. Pass 'full' instead of true for the entire page height.

Examples

"Rebuild stripe.com's pricing section in React"

The agent calls extract_code with the URL, gets the section's real markup, colors, fonts, and spacing, and adapts actual values instead of guessing.

Extraction returns a page's light styles by default, so a site you know as dark comes back light unless you ask otherwise:

"Rebuild linear.app's pricing card in its dark theme"

The agent passes theme: "dark" and the page's prefers-color-scheme: dark rules and custom properties become the component itself. Every result also carries a theme field saying which one you actually got. Sites that switch themes with a JavaScript class toggle cannot be selected this way - you get the page default plus a hint saying so.

"Build a dashboard in the style of linear.app"

The agent calls extract_brand and extract_design_tokens, gets Linear's real colors, fonts, and spacing scale, and builds to match.

"What fonts does apple.com use?"

The agent calls extract_fonts and returns every family with its font file URLs and weights.

Notes

• extract_code always loads the page in a real browser and reads what it actually shows (on every plan, no extra credits), so it never needs render: true. The other tools read static HTML and resolve the full CSS cascade first. When that comes back as a JavaScript app shell, with styles injected by scripts, or refused by the site (401, 403 or 406, or timed out - publishers commonly turn away datacenter IPs even for a public page), the server re-runs the call in a real browser by itself at no extra cost, on every plan, and the result carries usage.autoRender (a sentence to read to the user).
• Pass render: true to force the browser (+5 credits, any plan): for images, video or Lottie that only load on scroll, or when a result is still thin after the automatic pass. A 403 that says the browser was refused too means the site cannot be extracted. On extract_design_tokens, render: true only permits the browser: the plain read still runs first, and the 5 credits are charged only if the browser was actually needed.
• Every result carries usage with the credits left this month. If the account holds a credit pack (one-off credits that never expire, bought at https://miromiro.app/api/pricing), usage.packCredits is the pack balance left; packs are drawn after the monthly credits, so a free account with a pack keeps working past its 300 credits.
• Respect the sites you extract from: check their terms before reusing assets commercially.

Links

API docs · Playground (no account) · Pricing · Changelog · Connect page

MIT © MiroMiro