AI Compass / Knowledge cards
Have your own website checked from the outside
An agent that only reads the code doesn’t see what happens in the browser. REA (“Reverse Engineer Anything”) is an open-source tool with which an agent such as Claude Code opens a page in the browser and records what it loads and sends. It comes as an MCP server, an interface that gives the agent additional tools, together with a skill. The skill instructs the agent to tie every statement to evidence and to label it “seen”, “suspected” or “unknown”.
What it is for
You find errors, slow spots, a broken mobile view and exposed keys before others do. The labels show you which points in the report you can trust.
Steps
- Read the skill and the installation instructions before anything runs. REA puts a skill in the global Claude folder.
- Install it, according to the README with
npx rea-agents setup(Node.js 22.19 or later). In Claude Code, check with/mcpwhether REA is connected. - Write the audit brief: link to the page, what it is supposed to do, “check from the outside only, don’t read the code”, report with the labels seen / suspected / unknown.
- Have speed and layout checked, including at phone size.
- Have it check for exposed keys and passwords, with the report sorted by danger.
- For the code behind it, optionally use Cloudflare’s
security-auditskill. It works with several review agents and an independent cross-check. - Work through the findings and check again. What is only “suspected” counts as a bug only once there is evidence.
Limits
- REA only sees what arrives in the browser. The logic on the server stays hidden.
- The test pages in the video were deliberately prepared, for example with built-in keys. How much REA finds on a real site is open.
- Check only your own sites or ones you have permission to check. According to the README, REA is meant for software you own or are authorized to examine. The analysis of other people’s sites shown in the video doesn’t fall under that without checking the terms of use.
- The labels seen / suspected / unknown come from the video (2:11–2:43, one “suspected” case at 9:46). The skill file in the repo demands the same in other words: keep observations, inferences and unknowns apart. The README doesn’t name the scheme.
- According to the README, the Cloudflare skill runs third-party code only in an isolated environment (no network, fixed write paths). If that is missing, findings stay unconfirmed. According to the README, a single run found about half of what several runs found together.
- Both tools are third-party code with broad permissions (browser, files). Check them beforehand, see Related.
Links
- Vetting third-party skills: before installing REA and the audit skill.
- “Done” only with evidence: the label “seen” is evidence, “suspected” is not.
- Websites without the AI look: there, the phone check is part of the comparison with the design.
Sources
- AI LABS: GitHub’s #1 Trending Skill Is Insane (5 Use Cases), 15 Oct 2026, 1:38–4:20, 8:08–11:56 (prepared test pages 10:51–11:23). Sponsored by Firecrawl (6:31–7:22), plus self-promotion for AI Labs Pro
- Jack Roberts: 10 New Github Repos to Make Money in 19 mins, 12 Oct 2026, 12:19, 18:33–19:20. Several short links, not marked as advertising
- GitHub: morluto/rea (MIT, README and skill file
.agents/skills/reverse-engineer-anything/SKILL.md) and cloudflare/security-audit-skill (MIT), retrieved 16 Oct 2026
Status
Checked 16 Oct 2026 · Maturity new (tools available, not tested by me)
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab