你真的需要 REA 嗎？我測試了 Reverse Engineer Anything 與單獨使用 Claude Code 的對比

Do You Even Need REA? I Tested Reverse Engineer Anything vs Claude Code Alone

I gave Claude Code only the compiled production build of my app Shiftwire, with no source code, no source maps and no .git, and asked it to rebuild the whole thing. I ran it twice on the same build: once on its own and once with REA (Reverse Engineer Anything, rea-agents 6.4.0) connected as an MCP server. Both rebuilds were scored against the real source on 8 hidden business rules. Without REA: 8/8 in 40m 27s of Claude working time, $10.52 API-equivalent. With REA: 8/8 in 16m 46s, $12.45. Neither run flagged the overnight AM/PM bug in my app. One app and one run each, so treat the speed difference as a hint.

Two ideas worth stealing. Score a rebuild against rules hidden in the real code, like a setting that never fires or an opt-out list, because matching screens proves very little. And on Windows, REA's MCP server took 35.5 seconds to start cold while Claude Code waits 30, so start it once first or set MCP_TIMEOUT=120000 before you launch claude.

Only download REA from the official repo: https://github.com/morluto/rea Lookalike repos exist. Do not run them.

Tested on my own app. Not legal advice. Only analyze software you own or are authorized to examine.

I build AI systems like this for service businesses: voice agents, automations, internal tools. Book a call: https://thisisaqib.com

Chapters: 00:00 The test: rebuild my app from compiled code 00:31 What REA is 01:10 Fake REA repos 01:40 Setup on Windows, and what broke 02:41 The hidden rules 03:40 The prompt 03:50 Round 1: without REA 05:05 Round 2: with REA 06:22 The scorecard 07:45 What REA changed: speed and cost 08:53 Is this legal? 09:35 Verdict 10:13 Outro