replica-crew

Replica Crew

Everything for cloning software with coding agents, in one repo, for Claude Code and Codex. Web pages, web apps, desktop apps, Android apps, games, and original game mashups. One command takes a target from "I want my own version of X" to a tested, rebranded product, with 11 specialist subagents. First it scans your computer, read-only, and tells you which target types it can build. The repo bundles 15 more open-source clone, design-extraction and UI-to-code skills, and maps 80+ related tools.

ES: Todo para clonar software con agentes de código, en un solo repo, para Claude Code y Codex: páginas y apps web, apps de escritorio, apps Android, videojuegos y mezclas originales de videojuegos (p. ej. mecánicas de Minecraft con las de GTA V). Primero escanea tu computadora, sin tocar nada, y te dice qué puedes construir. 11 subagentes más 15 skills de clonado integradas. Sala limpia: se clonan funciones y mecánicas, nunca código, arte, marca ni contenido.

Claude Code:  /replica-crew https://example-app.com
Codex:        $replica-crew https://example-app.com

Five target types (web, desktop, android, game, mashup): see docs/TARGETS.md. To set up and scan the machine in one step, paste the prompt in docs/INSTALL_PROMPT.md. Check the machine any time: python skills/replica-capabilities/doctor.py.

Binary decompilation (Claude Code and Codex)

Both hosts support authorized static analysis through replica-recon. When replica/scope.md records ownership or written authorization, the authorized-analysis mode turns on and recon proceeds with the full toolbox: APK/DEX (JADX, Apktool), .NET (ILSpy), JAR/class (CFR), native PE/ELF/Mach-O and iOS IPA (Ghidra, Rizin), Flutter AOT (blutter), Unity IL2CPP (Il2CppDumper), Go (GoReSym), Electron ASAR (asar), Python bytecode (pycdc), or a connected MCP (GhidraMCP, ida-pro-mcp, jadx-ai-mcp, apktool-mcp-server, frida-mcp). The bundled bin_analyze.py inventories inputs, records SHA-256, detects format from magic bytes, and picks the tool. Recon writes replica/binary/report.md with evidence and a functional replica/behavior-spec.md for the builders; raw recovered code is kept separate and stays out of product-mode clones. Record ownership/authorization and analysis scope in replica/scope.md.

Claude Code: /replica-crew Analyze my local binary <path>; I own it. Static decompilation only.
Codex:       $replica-crew Analyze my local binary <path>; I own it. Static decompilation only.

See the shared binary workflow. Build tools are optional for analysis-only work; decompilers are not bundled or installed automatically.

Install

git clone https://github.com/damiantorresmx11/replica-crew
python replica-crew/scripts/install.py          # Claude Code and/or Codex
python replica-crew/scripts/install.py --mcp    # + Playwright and Chrome DevTools MCP

Or by hand:

host	command
Claude Code	/plugin marketplace add damiantorresmx11/replica-crew then /plugin install replica-crew@replica-crew
Codex	copy skills/* to ~/.agents/skills/ and codex/agents/*.toml to ~/.codex/agents/
Codex plugin marketplace	codex plugin marketplace add damiantorresmx11/replica-crew
any other agent	read AGENTS.md; every SKILL.md is plain Markdown

Requirements: Python 3.8+. Node for Playwright and some bundled skills. Tests: python -m unittest discover -s tests -v.

First run: install + scan your machine

Prefer a hands-off setup? Paste this whole block to Claude Code or Codex. It installs the toolkit, scans your computer read-only, and tells you which target types you can build before anything else runs. The agent installs only the toolkit; it never installs languages, SDKs or drivers for you. (Also kept at docs/INSTALL_PROMPT.md.)

You are setting up "Replica Crew" (https://github.com/damiantorresmx11/replica-crew),
a toolkit of skills and subagents for cloning apps, websites, desktop apps,
Android apps and games with coding agents. Do this in order and stop at the
first blocker.

STEP 1 - INSTALL THE TOOLKIT (this is the only thing you install).
- Detect whether you are Claude Code or Codex from your own tools.
- Claude Code:
    /plugin marketplace add damiantorresmx11/replica-crew
    /plugin install replica-crew@replica-crew
- Codex: clone the repo to a temp dir and run its installer:
    git clone --depth 1 https://github.com/damiantorresmx11/replica-crew
    python replica-crew/scripts/install.py
  (It copies skills to ~/.agents/skills and the 11 agents to ~/.codex/agents.)
- If git or python is missing, stop and tell me the one command to install it
  for my OS. Do not work around it.

STEP 2 - SCAN THIS COMPUTER (read-only, installs nothing, no network).
- Find doctor.py inside the installed "replica-capabilities" skill (look under
  ~/.agents/skills, ~/.claude/skills, ~/.claude/plugins, or the cloned repo's
  skills/ folder).
- Run:  python <path>/doctor.py --json
- The scan only runs "--version" probes and checks a few folders. It reads my
  OS, CPU count, and which toolchains I have: Node, npm, Python, Git, Docker,
  Java, Gradle, Flutter, Android SDK, .NET, Rust (cargo), Go, CMake, Godot,
  Blender, FFmpeg, Playwright, adb, gh. It writes nothing except its own report.
- If you cannot find or run doctor.py, say so and stop; do not guess my setup.

STEP 3 - TELL ME WHAT I CAN BUILD.
From the JSON "targets" block, give me a short table of the five target types
(web, desktop, android, game, mashup) with READY or NOT-READY for each, and for
every NOT-READY one, the exact tools to install and their official sites (the
scan already names them). Then one line: "Ready now: <list>."

STEP 4 - ASK, DON'T ASSUME.
Ask me which target and which app/site/game I want to clone. Do not start recon
or the build until I answer. If I pick a NOT-READY target, list what to install,
let me install it, then re-run STEP 2 - never install it for me, never fake the
scan, and never promise a stack my machine can't build, test or run.

STEP 5 - RULES YOU KEEP THE WHOLE TIME.
- Clean room: rebuild features, flows and mechanics from public sources and my
  own account or my owned copy. Never copy the original's code, assets, art,
  audio, trademarks or content, and never decompile a binary or load/mod a game
  I do not own unless I state in writing that I own it or am authorized.
- Never create accounts, buy domains, type passwords, or paste live keys.
- Nothing ships under the original's name, and nothing goes live without my go.

Report each step's result in one or two lines. Begin with STEP 1.

Three modes

mode	for	engine
product (default)	your own product with the features of someone else's app	11 Replica agents, clean room
faithful	a site you own or are authorized to copy: migrate your site, restore a lost codebase, client work under contract	clone-team, clone-website, web-clone, stitch-extract-static-html; faithful-capture for a verbatim copy that keeps the real animations
study	learn how a premium site is built, kept local, assets substituted	clone-study

Product mode: how it runs

Wave -1 capabilities         read-only machine scan: which targets are buildable
Wave 0  scope (you)          target type, app, slice, stack, how far, ownership
Wave 1  recon                screens, flows, data model, feature matrix
Wave 2  architect | design | entrepreneur     <- in parallel
Wave 3  build -> backend
Wave 4  test -> diff -> build gaps            <- loop, max 3 rounds
Wave 5  brand -> launch
Wave 6  deploy preflight -> your explicit go -> live

• Specialists. Each phase runs in its own subagent (Claude: replica-crew:replica-X; Codex: custom agent replica-X), so the main conversation keeps only short reports.
• Parallel wave. Architect, design and the review-reading entrepreneur run at once; the fixes users ask for enter the feature matrix before the build.
• Quality loop. Test and diff repeat on the gap list until the verdict is shippable (feature score 80+, no open S1/S2), max three rounds.
• Two stops only. After recon and before go-live. Agents that need a human decision return NEEDS_USER instead of guessing.
• Audit trail. Every agent appends to replica/crew-log.md.

The eleven agents

Agent	Does	Writes
replica-recon	maps the app from public sources and your own account	recon.md, features.csv, screens/
replica-architect	stack, SQL schema, routes, build order	architecture.md, schema.sql
replica-design	design tokens and component specs, WCAG checked	design/
replica-entrepreneur	reads real, linked user reviews; fix plan and angle	feedback.md, fixes.md
replica-build	screen by screen, every state	app code, build-log.md
replica-backend	auth, DB rules, Stripe test mode, email, jobs	migrations, backend.md
replica-test	test plan, Playwright, bug log by severity	test-plan.md, bugs.md, e2e/
replica-diff	feature parity score and layout diff	parity.md, diffs/
replica-brand	new name, palette, logo brief, voice, leftover sweep	brand.md, brand.json
replica-launch	landing, pricing, store listing lint	launch/
replica-deploy	preflight gate, DNS, production, monitoring	deploy.md

Seven standard-library Python tools ship with them: bin_analyze.py, imgdiff.py, parity.py, reviews.py, contrast.py, sweep.py, listing.py.

Bundled skills (15 third-party, all MIT or Apache-2.0)

skill	from	used in
clone-team	Varalix-Digitech-Solutions/clone-team	faithful: whole sites with a 4-agent team
clone-website	JCodesMore/ai-website-cloner-template	faithful: pages to Next.js fast
web-clone	Jane-xiaoer/claude-skill-web-clone	faithful: WebGL, Canvas, Three.js
clone-study	luukalleman/clone-study	study mode
design-mirror	brightdata/skills	apply a look to your existing app (Bright Data key)
stitch-extract-design-md, stitch-extract-static-html, design-md, site-md, stitch-react-components, shadcn-ui	google-labs-code/stitch-skills	recon, design, build
video-to-ui	mmohajer9/video-to-ui	recon from walkthrough videos
ui-image-to-code	zwq-top/ui-image-to-code-studio	build from your own wireframes
restyle	Mahanaicoach/ai-site-cloner	brand
app-store-screenshots	ParthJadhav/app-store-screenshots	launch

Each keeps its license (LICENSE.upstream) and source commit (UPSTREAM.md). Details: THIRD_PARTY_NOTICES.md.

Ecosystem

docs/ECOSYSTEM.md maps 70+ tools verified on GitHub in October 2026: REA, screenshot-to-code, open-lovable, Chrome DevTools and Playwright MCP, Figma MCPs, dembrandt, design-extract, DESIGN.md, BackstopJS, pixelmatch, GhidraMCP, jadx and apktool MCPs, mitmproxy2swagger, crawl4ai and more, with the phase and agent each one fits. Apps, libraries and MCP servers are linked rather than copied: they are separate products with their own runtimes and release cycles.

Rules

• Product mode rebuilds what an app does, never what it owns: no source code, bundles, private APIs, logos, icons, licensed fonts, copy or content.
• Public sources and your own account only. No logging into other accounts, no scraping behind logins, no hammering their servers.
• Faithful mode and binary analysis apply to software you own or are authorized to copy or analyze, recorded in replica/scope.md; once recorded, the full toolbox is available.
• No fabricated reviews, testimonials, numbers or "available" checks.
• Nothing launches under the original's name, and nothing goes live without your go.

Run the trademark checks replica-brand lists before you charge money.

Credits and license

MIT. Replica skills and tools © 2026 Jake Schincariol (Jakeschincariol/replica-skill, upstream README). Bundled skills © their authors under their own licenses. Agents, orchestrator, installer and docs © 2026 Damian Torres / DMLABS.