网站克隆工具

📦 Dependencies

These install automatically — clone-team ships scripts/install-deps.sh, an idempotent installer it runs on first use (and which you can run yourself any time: bash scripts/install-deps.sh, or --check for a dry run). By default it installs the companion skills project-local — into ./.claude/skills of the folder you're cloning in — so it never pollutes your global ~/.claude/skills (pass --global if you do want them installed globally). It skips whatever's already present:

• ui-pack — design/frontend skill bundle (the single entry point the agents load); vendored with this plugin, so it's always available. It in turn loads its constituents below.
• clone-website — extraction + builder dispatch
• ui-ux-pro-max, impeccable, emil-design-eng — design intelligence + polish
• ui-animation — motion craft (transitions/keyframes/springs, easing, clip-path reveals, gestures, performance) for the two motion specialists
• agent-browser — real-browser automation (npm CLI) for building and verifying; the one hard dependency (needs Node/npm on PATH)

Every piece degrades gracefully if absent except agent-browser, which is required for real-browser verification.

Clone any website with a coordinated team of AI agents — and get the architecture docs to rebuild it.

A Claude Code skill that orchestrates a Manager, Frontend Developer, Backend Architect, and Tester to produce a pixel-perfect UI clone and a reverse-engineered ARCHITECTURE.md — in the stack you choose.

🌐 Website, demos & docs → clone-team.varalix.com

If this saves you a week of pixel-pushing, drop a ⭐ — it genuinely helps.

<br/>

<img src="docs/media/wembi-compare.gif" alt="wembi.ai — original (left) vs clone-team build (right), scroll-synced side by side" width="100%">

_{<b>Original (left)</b> vs. <b>clone-team build (right)</b> — <a href="https://www.wembi.ai/">wembi.ai</a>, driven in lockstep so both stay at the same scroll point. A hard target: Lenis smooth-scroll + scroll-driven 3D + reveal animations. <a href="case-study/wembi/">Full case study →</a>}

</div>

---

🚀 Install

Recommended — install as a plugin (one marketplace, one install, auto-updates). In Claude Code:

/plugin marketplace add Varalix-Digitech-Solutions/clone-team
/plugin install clone-team@clone-team

That's it — the skill and the /clone-status, /clone-pause, /clone-resume, /clone-report, and /clone-update commands are registered automatically. Plugin installs auto-update when a new version ships; run /clone-update anytime to check what you're on vs. the latest. (Restart Claude Code if the commands don't show up immediately.)

You don't install the companion skills yourself. The first time you run a clone, clone-team bootstraps its own toolchain — it runs a bundled, idempotent installer that fetches the agent-browser CLI and the companion skills (ui-pack and friends, ui-animation) project-local (into ./.claude/skills, so your global skills stay clean), skipping anything already present. Just install the plugin and ask it to clone a site; it sets itself up and then gets to work. See Dependencies.

<details> <summary><b>Alternative — manual skill install (no plugins)</b></summary>

```bash git clone https://github.com/Varalix-Digitech-Solutions/clone-team.git cp -r clone-team/skills/clone-team ~/.claude/skills/clone-team

🧑‍💻 How to use it

Just ask Claude Code to clone a site — natural language triggers the skill:

Clone https://example.com — the marketing homepage — in Next.js,
and document how the site is put together.

The Manager then walks you through a short setup (you stay in control):

• Stack — what to build the clone in (React / Next.js / Vue / Svelte / plain HTML…).
• Scope — whole site, or specific pages.
• Login — if the target needs auth, you provide credentials; they're stored in a gitignored local file and never printed, committed, or pasted into prompts.
• Model tier — max-fidelity (Opus, default), cost-optimized (Sonnet + Haiku), or ultra-cheap.
• Autonomy — fully autonomous, or checkpoint at each Tester verdict for sign-off.
• Backend depth — none / flows / inferred / deep.

Then it runs the dynamic Workflow in the background and reports back with the clone + ARCHITECTURE.md.

optional: the slash commands

cp clone-team/commands/clone-*.md ~/.claude/commands/

bootstrap the companion skills + agent-browser CLI (idempotent)

bash ~/.claude/skills/clone-team/scripts/install-deps.sh ``` </details>

Requirements: Claude Code, Node.js + npm (for the durable state CLI and to install agent-browser), and a Chromium browser (for real-browser verification via agent-browser). Everything else installs automatically — see Dependencies.

⚙️ The engine: a **dynamic Workflow**

clone-team is built on Claude Code's new dynamic Workflow engine — and it's the whole reason the quality guarantee holds.

A dynamic Workflow is a deterministic JavaScript program that spawns and coordinates a fleet of subagents — where the loop count, fan-out width, and exit conditions are all computed at runtime from real data, not hardcoded. clone-team uses it so the process can't drift:

• The Tester gate is unskippable. Per page: extract → spec → build → full-regression-test → fix → re-test … runs as a while loop whose exit condition is the Tester's OK. No human, no model, and no "looks fine under deadline pressure" can bypass it.
• It adapts to the work. One-page site → one builder. Nine-page site → nine, running concurrently. The Workflow reads the page list and fans out accordingly.
• It sizes the agent fleet to your PC. Before launching, a capacity probe reads your machine's free memory and computes how many builder agents to run at once — many in parallel on a big box, a few at a time on a laptop — so it goes as fast as your hardware allows without ever OOM-crashing it.
• It feeds failure back as input. A Tester NG isn't just a fail; its structured issue list becomes the next Developer round's fix list.
• It's pausable & resumable — even across sessions and usage-limit cutoffs — because progress is reconciled against a durable state.json + on-disk artifacts.

user ⇄ MANAGER (main thread)         requirements, creds, recon, foundation,
        │                            final regression, pause/resume
        │ launches a DYNAMIC WORKFLOW that ENFORCES the loop
        ▼
   per page, in parallel:   extract → spec → DEVELOPER builds →
                            TESTER full regression → (NG) fix → re-test → OK ✅
   in parallel:             BACKEND ARCHITECT writes ARCHITECTURE.md
   then:                    assemble → final regression → fix

Two gates protect quality — the Tester (inside the loop) and the Manager (final sign-off) — and nothing ships past both.