Today, a terminal AI agent needs just 3–5 minutes to dissect an unfamiliar web app from a single public URL and build a working local prototype without server access. It feels like without source code you could only grab a silent shell of images and buttons from someone else's site, and manually inspecting requests and minified scripts in DevTools would eat half a day.
The open-source utility reverse-engineer-anything (REA) automates this reconnaissance for Claude Code, Cursor, and Windsurf. In automation practitioner Julian Goldie's breakdown, one detail stood out to me: he gave Claude the link to his consultation booking page, and the model extracted call durations, buffer pauses between slots, and manager names on its own. How does an algorithm learn a service's internal rules from an external page—and where does that kind of assembly hit a wall?
Six Steps from a Browser Link to a Local Prototype
An interactive website works like a coffee shop pickup counter: to understand order logic, you don't need to break into accounting—the register and the barista are already exchanging tickets in plain view of the customer. As long as the page is live, the server itself sends data packets and field-validation scripts to the browser, and REA connects to the agent via the Model Context Protocol (MCP)—an adapter into the browser inspector. Suppose you are building your own online booking service modeled after Calendly without paying for third-party subscriptions; the free open-source package installs in 1 minute:
npx rea-agents@latest setup
The command registers the MCP server in your agent's settings (from Claude Code to the Hermes environment), after which you simply paste the calendar URL into the chat. A strict chain kicks in—once traffic is intercepted, the response structure is known, and the interface is rebuilt from it:
- Network interception: a script simulates clicks on dates and captures
XHRandFetchrequests along with their data payloads. - DOM tree parsing: the parser strips framework wrappers, isolating the slot grid, time zones, and form fields.
- Script deobfuscation: static analysis of minified JS bundles (and
ASARarchives in Electron) recovers function names, validation rules, and buffers between calls. - Data schema recovery: strict TypeScript interfaces are automatically generated from the server's JSON responses.
- Component synthesis: the agent rewrites the frontend cleanly as React or Next.js modules styled with Tailwind CSS.
- Local prototype assembly: mock Node.js API handlers are created to run standalone on your computer.
A live app appears in your working folder, matching the original's responsive grid with up to 90% accuracy. In Goldie's test, the calendar immediately reproduced the schedule, business hours, list of consultants, and guest link: the initial layout remains a draft for manual polish, while the model handled date states and form-to-API wiring on its own.
Visual Clone vs. Closed Backend Logic
When the local calendar briskly clicks through slots, it is easy to mistake the storefront for a finished product. Roll out a prototype with a mock backend to real customers—if two people book Tuesday at the same time, the stub will show success to both, while neither conflict checks against live team calendars nor lead routing will work. I checked the morluto/rea repository documentation: the parser sees only the interface and external API signatures, whereas the closed server—internal database transactions and ranking algorithms—stays behind the wall. The agent traces the shape of the wall outlet, but you are the one wiring up the power plant.
Keep two practical limits in mind as well. With heavy dynamic minification, a human will need to label event handlers manually, and WebAssembly modules require decompilers—as covered in our breakdown of pairing REA with the Ghidra disassembler. Directly copying someone else's logos, graphics, and code violates copyright and Terms of Service, so mechanics are transferred via independent implementation, much like the clean-room method for free alternatives to paid software. Used strictly as a scaffolding generator, however, REA speeds up calendar assembly many times over: you get clean React modules and TypeScript contracts, swap in your own design, and hook up your own database.
Don't start with massive third-party portals—take your own old lead form or call-booking landing page today. Connect rea-agents in the terminal, give the agent the link, and ask it to dump the discovered field-validation rules and data schema into a text file—a quick X-ray will immediately show which part of the project is ready for a fresh stack.

