dittu
The dittu MCP server lets an agent clone, verify, and inspect website frontends generated from live rendered pages.
Clone a website frontend via
clone_website: capture markup, original CSS, fonts, images, and scroll animations into a static preview plus Next.js and Astro projects.Crawl multiple pages with
pages(up to 50), and remove unwanted elements with CSS selectors viaremove.Choose output styling: keep original CSS by default or convert losslessly to Tailwind utilities with
styling: "tailwind".Build and score generated apps with
build: true, producing pixel-diff match scores at desktop and mobile widths.Verify an existing clone via
verify_clone, especially after refactoring generated components, acrosspreview,next, andastrotargets and custom widths.Read clone reports via
read_report: pages, targets, per-width scores, diff image paths, and worst regions.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@dittu Clone https://example.com into Next.js and verify the visual match."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
dittu: Clone Any Website
Turn any website into an editable Next.js or Astro project, verified pixel by pixel.
Layout, CSS, fonts, images and scroll animations. Not the backend. One CLI, one MCP server, and skills for your coding agent.
Quick start · How it works · Benchmarks · Docs
$ dittu clone https://gole.codes --build
› Capturing https://gole.codes
1,680,755 chars of markup, 6 stylesheets, 33 animations, 13 JS-responsive elements
✓ Static preview → dittu-out/gole.codes/preview
✓ Next.js project → dittu-out/gole.codes/next
✓ Astro project → dittu-out/gole.codes/astro
Visual match (1 − differing pixels / total)
preview /home 1440px 99.95%
next /home 1440px 99.95%
astro /home 1440px 99.95%
preview /home 390px 99.89%
next /home 390px 99.89%
astro /home 390px 99.89%
✓ All checks ≥ 95% (lowest 99.89%)Quick start
Set up your agent
npx @manishtamang/dittu setupRegisters the dittu MCP server and installs the skills for every agent it
finds: Claude Code, Cursor, Gemini CLI, Codex. Check your machine first
with npx @manishtamang/dittu doctor (on Linux/WSL you may need
sudo npx playwright install-deps chromium).
Ask your agent
Clone https://gole.codes into Next.js and make the pricing cards a reusable component.
The agent calls clone_website, reads the scores, refactors the generated
components, and calls verify_clone to prove nothing visual changed.
Use the terminal
npx @manishtamang/dittu # interactive: banner, then pick URL, frameworks, pages…
npx @manishtamang/dittu clone https://gole.codes -y # no questions: preview + Next.js + Astro, scored
npx @manishtamang/dittu clone https://gole.codes -p 5 --build # crawl 5 pages, build and score the apps
npx @manishtamang/dittu clone https://gole.codes -s tailwind # Tailwind classes where lossless, rest stays CSS
npx @manishtamang/dittu verify dittu-out/gole.codes --targets next
npx @manishtamang/dittu serve dittu-out/gole.codesUpdate dittu
npx @manishtamang/dittu@latest setupRelated MCP server: PicassoWeb
How dittu works
Most cloners screenshot a page and ask an LLM to redraw it. dittu copies what the browser actually rendered, then checks the result:
Capture: Chromium loads the page, scrolls it so lazy content and scroll reveals fire, then freezes JS timers so carousels hold still. A second render at 390 px records layout that JS computes per viewport.
Serialize the live DOM, not the server HTML. Next.js/React sites come out as their final markup without hydration scripts or RSC payloads. Nesting the HTML parser would restructure gets repaired, so React hydrates cleanly.
Keep the original CSS in cascade order, including CSS-in-JS rules that exist only in the CSSOM. Media queries, hover states and keyframes keep working.
Download every asset from the same session and rewrite URLs.
Export to a static preview, a Next.js App Router project (JSX) and an Astro project, split into section components.
Replay motion: recorded scroll reveals, Framer Motion
animatecalls and loops run again through a 2 KB runtime.Verify: screenshot original and clone the same way at desktop and mobile and score them.
--buildscores the built apps, not just the preview.
Details: docs/architecture.md.
What you can clone
Site type | Support |
Marketing sites, landing pages, portfolios, docs | ✅ 99%+ typical |
Next.js / React / Vue / Svelte / Astro builds | ✅ post-hydration DOM |
Tailwind, CSS modules, styled-components, emotion | ✅ original CSS kept |
Scroll reveals (AOS, Framer Motion, GSAP, IntersectionObserver) | ✅ recorded and replayed |
Layout JS computes per viewport | ✅ mobile render → media-query overrides |
Multi-page sites | ✅ |
Canvas / WebGL scenes | ⚠️ still image |
Carousels, tabs, menus, forms | ⚠️ look is copied, behaviour is not |
Pages behind login or bot protection | ❌ |
Benchmarks
Site | Stack | Desktop | Mobile | Built apps |
gole.codes | Next.js, heavy animation | 99.95% | 99.89% | Next.js + Astro: same |
nextjs.org | Next.js | 99.79% | 99.99% | 99.78–99.79% |
ui.shadcn.com | Next.js + Tailwind | 100.00% | 100.00% | 100.00% |
astro.build | Astro | 100.00% | 100.00% | |
manishtamang.com (3 pages) | Next.js | 99.89–99.99% | 99.80–100% | same |
Methodology and limits: docs/fidelity.md.
FAQ
Does it copy the backend? No. Data fetching, auth, APIs and form handling are not copied. You get the frontend, ready to wire to your own logic.
Why is the generated JSX full of {" "}? Whitespace between inline elements
changes layout in HTML but is dropped by JSX. dittu keeps it explicit so the clone
renders identically. Refactor freely, and re-run verify to check.
Can I get Tailwind classes instead of CSS? Yes: --styling tailwind. Every
rule that can be expressed exactly becomes Tailwind utilities (sites built with
Tailwind v4 keep their own classes); the rest stays CSS, so scores don't drop.
See Tailwind output.
Is the 95% score honest? It's raw pixel identity over the full page, including height differences, at desktop and mobile. Diff images are saved next to the scores so you can check.
Can I clone any site I want? Technically mostly yes. Legally, see the disclaimer.
Documentation
Disclaimer
dittu is for learning, prototyping, design references and migrating sites you own. Cloning a site doesn't give you rights to its brand, copy, images or code. Read DISCLAIMER.md.
Related projects
ditto.site by ion-design takes the same capture-what-the-browser-rendered approach as a hosted API and MCP service. dittu differs by running as a local CLI, keeping the site's original CSS by default (Tailwind output is opt-in and only where it renders identically), exporting Astro alongside Next.js, and scoring every clone with pixel-diff verification.
Contributing
See CONTRIBUTING.md. Agents working in this repo: AGENTS.md.
License
Available Tools
3 toolsclone_websiteClone website frontendA
Clone a website's frontend (markup, original CSS, fonts, images, scroll animations) into a static preview plus Next.js and Astro projects, then score it against the original with pixel diffs at desktop and mobile widths. Takes ~1 minute per page (several minutes with build=true). Returns output paths and scores (percent identical).
Inputs: url (required), build, pages, outDir, +5 more.
read_reportRead clone reportBRead-only
Manifest and latest scores of a clone: pages, targets, per-width scores, diff image paths, worst regions.
Inputs: outDir (required).
verify_cloneVerify a cloneA
Re-score an existing clone against the original site (e.g. after refactoring the generated components). Use targets ["next"] or ["astro"] to build and score a generated project. A score drop means the refactor changed rendering.
Inputs: outDir (required), widths, targets.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
clone_website and verify_clone both produce scores, which is a slight overlap, but the descriptions clearly separate creating a new clone from re-scoring an existing one. read_report is cleanly distinct as a read-only manifest lookup.
All three tools follow a consistent verb_noun snake_case pattern: clone_website, read_report, verify_clone. Predictable and easy to scan.
Three tools map precisely onto the create/read/verify workflow of a cloning service. Nothing feels padded and the scope is tight for a specialized pipeline.
The core lifecycle (create clone, inspect report, re-verify after refactor) is fully covered. Minor gaps exist around housekeeping, such as listing or deleting existing clones, but agents can work around these via paths.
Maintenance
Related MCP Connectors
Real HTML, styles, measurements and brand tokens from any live site, for coding agents.
Captured web interfaces, screenshots, flows and structured evidence for AI coding agents.
Website checker for AI coding agents. Measures your site in a real browser, with evidence.
Give your agent eyes: screenshots of any web page it can see, pages as markdown, and PDFs.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI coding agents to visually interact with frontend apps by taking screenshots, clicking elements, reading console logs, and performing visual diffs.6 npm3MIT
- AlicenseAqualityCmaintenanceAnalyzes and extracts design tokens, assets, and layout from live websites to enable AI clients to faithfully replicate them, with tools for screenshotting, component inspection, and pixel-diff verification.1418 npm4MIT
- AlicenseNot gradedqualityCmaintenanceEnables copying any component from a live web page as clean, self-contained HTML or JSX from inside a coding agent. Renders pages in a real browser to capture computed styles and resolved layout, with configurable styling formats and caching.MIT
- AlicenseAqualityBmaintenanceEnables AI coding agents to see, measure, and verify web pages through a real Chrome browser, including screenshots, responsive layout and accessibility audits, pixel diffing against baselines, secure logins, and deployed-fix verification.271MIT