REA Reverse Engineer Anything GitHub: The Official Repo, How It Works, And The Models That Can Drive It

REA Reverse Engineer Anything GitHub is the search, and github.com/morluto/rea is the answer, because that is the official repo and it is free and open source under the MIT licence.

You install it with one command, which is npx rea-agents setup.

Check that the repo owner is morluto before you run anything, and never download a zip installer from a lookalike repo, because the real project installs through npm.

REA stands for Reverse Engineer Anything, and it is one MCP server that lets an AI agent inspect apps, websites and binaries and explain how a feature works.

I'm Julian Goldie, and I tested it with Claude on my own booking page.

In this post I'll cover what is in the repo, how REA works, how to set it up, what it can analyse, my real test and the honest limits, and then I'll look at the models that can drive it.

What is in the REA Reverse Engineer Anything GitHub repo?

These are the facts I checked on the repo, the README and the npm registry on 13 October 2026.

The star count moves every day, and my video showed 181,000 when I recorded it.

What REA is and the problem it solves

REA is a set of tools your AI agent can call when you want to understand how a piece of software works.

The problem is simple.

You see a feature in an app or on a website, you want something like it in your own project, and you have no source code to learn from.

Without REA, an agent can only guess from what you tell it.

With REA, the agent inspects the target, traces the relevant code and returns findings with the evidence behind them.

The README's example prompt is "Understand how search works in the Notes app, show me the evidence, and build a similar feature for my project."

You replace Notes with your own target and the feature you want to understand.

How REA works

Your agent calls REA through MCP to inspect the target and trace the relevant code.

REA returns findings with their evidence.

Your agent then asks follow-up questions, explains the behaviour, or writes and tests an implementation.

The README says results include the evidence and the limitations behind each conclusion.

On privacy, the README says REA analyses targets locally, your agent receives the tool results, and your model provider has its own data policy.

The same workflows also run from the terminal.

How to check you have the real REA GitHub repo

Search results for REA GitHub include copycat repos that are stuffed with keywords and offer an "official free download" as a Windows zip.

The real project is not a zip download.

It installs through npm with npx rea-agents setup, and it lives at github.com/morluto/rea.

The owner in the address bar should be morluto, and the README should show the setup command near the top.

You should close any page that pushes a zip or an installer file at you.

How to set up REA from GitHub

The README asks for Node.js 22.19 or newer, 24.11 or newer, or 26 and above, plus npm.

With those installed, you run this command.

npx rea-agents setup

You choose your agents, review the proposed changes and approve them.

Setup adds REA's MCP server and matching workflow instructions, and it backs up your existing configuration.

You restart your agent afterwards.

Setup supports Claude Code, Codex, Cursor, Gemini CLI, Grok Build and other agents.

The README says any agent that supports local MCP servers can use REA through manual MCP registration.

You update a global install with rea update, and you refresh an npx setup with npx rea-agents@latest setup.

What REA can analyse

The README lists a lot of targets, and each one has its own requirements.

The README also lists WASM files, EVM bytecode, SQLite databases, Go binaries, packet captures and process behaviour.

Several of those are Linux-only or need tools you supply yourself.

Setup can optionally install Hopper with your approval, while Ghidra and IDA use the copies you already have.

Using REA from the terminal

The README shows this command for an extracted JavaScript or Electron app folder.

npx -y rea-agents@latest analyze-javascript-application /absolute/path/to/app --json

The result includes modules, imports, Electron boundaries and their evidence.

You can also install the command for regular use.

npm install --global rea-agents
rea --help

My real test: Claude and my own booking page

This is the only test I am claiming.

I used Claude inside my Agent OS.

I gave it the REA GitHub link and a link to one of my own funnels, which is the booking page for my free SEO strategy session.

I gave it no back-end access, so it worked from the front end only.

Claude found the scheduling rules, including appointments per slot, slot duration, slot interval and calendar type.

It found who on my sales team is on the calendar, and it worked out the calendar API.

Then it built my own custom scheduler app and plugged it into the workspace in my Agent OS.

Claude also built a page explaining how reverse engineering works.

It took a couple of prompts.

I think we could improve the design, and I would change the UI.

I am not a technical person, which is why this test mattered to me.

If you want to see the workspace it lives in, read my Agent OS GitHub post.

The six steps of reverse engineering

Claude explained the idea with two analogies, and I am crediting Claude for both.

One is tasting a sauce at a restaurant and working out the recipe, and the other is watching someone's golf swing.

It then gave me six steps.

The REA showcases

The project site lists three case studies.

DX-Ball reconstructs a sound-pan calculation, and the README says it passes 3,205 original-x86 cases.

Notion traces the Electron clipboard bridge from the renderer through preload and IPC into the main process.

TH04 recovers a PC-98 DOS game's bullet-ring calculation.

The Notion case study traces one clipboard feature, and it is not a clone of Notion.

Limits and honest caveats

I did not test native binary analysis, Hopper, Ghidra, IDA, Android or firmware.

Those are documented features, and I am reporting what the README says about them.

I only tested REA with Claude.

In the video I also pasted the GitHub link into Hermes, and I said the skill can be taught to agents like Antigravity and Muse, but I did not show results from them.

The scheduler works, but the design needs improving.

The README says REA changes quickly and ships frequent bug fixes, so you should expect to update often.

Commercial tools such as IDA and Hopper need their own legal licences, and REA does not provide one.

Use REA responsibly

Use REA on things you own or are allowed to examine.

The README says it is for lawful research, education, interoperability, software preservation, and analysis of software you own or are authorised to examine.

Learn how a feature works, and then build your own original version.

Do not copy someone else's code, brand or content.

REA doesn't ship its own model, so which model should drive it?

REA does not come with an AI model.

It is a set of tools, and the model inside your agent does the reasoning.

That means the agent you connect decides how good the investigation and the rebuild will be.

I used Claude for my test, so I'll start with the Claude models on my leaderboard.

I want to be clear about what GoldieBench can and cannot tell you first.

GoldieBench gives every model the same one-shot prompt to build a game, a page, a simulation or a visual, and then I score the result from 0 to 10.

REA asks a model to do something different, which is to call tools over a long session, read evidence and then build.

So a GoldieBench score measures raw building ability, and it is not a score for how well a model runs REA.

Nobody has benched REA itself on the board, and I won't pretend otherwise.

The scores are most useful for the last step of reverse engineering, which is rebuilding the feature as your own version.

The Claude models on GoldieBench

I used Claude in the video, and these are the Claude models scored on the board.

Which Claude model your agent actually uses depends on your plan and your settings, so check that before you compare.

Other models an MCP agent could run with REA

REA's setup supports Codex, Cursor, Gemini CLI and Grok Build as well as Claude Code.

I have not tested REA with any of them, so this table only shows how models from those makers score on one-shot builds.

The model your own agent runs depends on how you have configured it.

If you want to run an agent on a free local model, the local models board shows how those options score, and I have not tested REA with a local model.

How I would choose a model for REA

I would start with the agent and the plan you already pay for, because REA itself costs nothing.

If you are choosing from scratch and you care most about the rebuild, the board says Claude Opus 5 and GPT-5.6 Sol are the strongest builders of the models above.

If you want a second opinion on a finding, you can ask a different model to check the evidence, because REA returns the evidence with each conclusion.

I would not pick a model on score alone, because long tool-using sessions are a different skill from one-shot builds.

If you are comparing agent setups, my StarNet GitHub post and my Hermes Agent OS GitHub post match other free repos to the same scores.

Also On Our Network

• Agent Operatorsthe step-by-step REA setup from GitHub for operators
• AI Income Deskseven honest ways a business can use REA
• AI Tool Verdictevery REA feature reviewed, with scores for what I tested
• agentos.guidethe REA quick start and where it fits in an Agent OS
• agentos.guidemy full Reverse Engineer Anything guide

Check the owner is morluto, run the setup command, and connect the model you trust for the rebuild, and you will get the most out of the REA Reverse Engineer Anything GitHub repo.