REA: Turn Claude Code Into a Reverse Engineer (MCP Guide)

Last updated: October 12, 2026

Most coding agents can only see code you already have. REA (Reverse Engineer Anything) flips that: it wires your coding agent directly into professional reverse-engineering tools like Ghidra, Hopper and IDA, so it can investigate compiled software you have no source code for — and show its evidence. In the week after it launched (7 October 2026), the open-source project picked up tens of thousands of GitHub stars, hitting the top of GitHub’s trending page (Startup Fortune, 9 October 2026). This guide walks through what REA does, the 15-minute setup, three safe first tasks, and — this part matters more than the install — the rules that keep you on the right side of the law and your own laptop’s safety.

Quick answer: REA is a free MCP server plugging Claude Code, Cursor and Codex into Ghidra, Hopper and IDA. Setup steps, three safe first tasks, and the rules for using it legally.

What REA is — and what it is not

REA is an open-source MCP server plus a command-line tool, maintained by a developer going by morluto (repo: github.com/morluto/rea; if you are new to the protocol, our MCP Explained guide covers how these servers work). Its pitch is a three-step loop: decompile a target to recover readable code and naming clues, trace execution until a feature is explained, then recreate the logic in your own stack (Startup Fortune, 9 October 2026).

It is not a source-code thief button. The project’s own motto is understanding mechanisms, not imitating features, and its documentation says every conclusion comes with evidence and stated caveats attached — the agent must tell you what it found, how confident it is, and what it does not know (agentbreaking.com, 12 October 2026). Analysis runs on your own machine rather than a hosted service, so the binary you inspect never leaves your hardware — though the decompiled code it reads does go to whatever AI model your agent uses, which matters for the safety rules below.

What it can analyze

One installer covers a spread that used to need several separate tools:

Native binaries via Hopper, Ghidra, or — since version 4.1 — IDA Pro (read-only and headless providers).
JavaScript and Electron apps, including static analysis that does not require a native backend.
.NET assemblies and websites.
Android APKs via headless JADX and firmware images via Binwalk and Unblob (both added in v4.1, released 6 October 2026).
Observed process behavior and runtime capture.

It talks to Claude Code, Codex, Cursor, Gemini CLI and Grok Build, among other MCP-capable agents, and earlier one-engine projects already did this niche: GhidraMCP (LaurieWired, March 2025) and ida-pro-mcp (mrexodia, March 2025) wired agents into single disassemblers. REA’s difference is breadth and packaging — many engines and formats behind one installer, plus evidence-and-unknowns output on top (agentbreaking.com, 12 October 2026).

Setup: 15 minutes, step by step

The honest caveat first: this guide is desk-researched, not hands-on tested. Steps below follow the project’s official setup guide and corroborating coverage, but run them in a throwaway environment the first time.

Meet the prerequisites. Node.js and npm are required. For native-binary analysis you will also want one of Hopper, Ghidra or IDA installed (static JavaScript analysis needs no native backend). Supported agents include Claude Code, Codex, Cursor, Gemini CLI and Grok Build (singulism.com, 10 October 2026).
Run the installer. The documented command is npx rea-agents setup. On the setup screen you select the target agents and review the proposed configuration diff.
Review and approve the diff. Existing agent settings are backed up first, and changes are applied only after your approval. Read what it wants to change before accepting — this is the single most important step.
Restart your agents. Agents must be restarted after setup for the MCP server and playbooks to register.
Pick a safe first target. Your own compiled program, an open-source project’s binary, or a CTF challenge file. Ask the agent a narrow question (“Which function handles the save action in this app, and what does it call?”), then check the evidence bundle it returns. If this is your first autonomous agent, our OpenPilot guide covers the same approval-first habits for desktop agents.

Three first tasks to try

Try this: “Decompile this old utility of mine and list the three features it implements, with the function names as evidence.” What happens: the agent decompiles, names clues from recovered code, and records provenance instead of guessing.
Try this: “This Electron app crashes when I click export — trace the export feature and explain what goes wrong, flagging anything you are unsure about.” What happens: static and dynamic analysis are correlated; gaps are stated as unknowns.
Try this: “Explain how this open-source library’s parsing logic works, then show me how I would implement the same behavior in my own code.” What happens: the “understand, then recreate” loop — the licensed-safe end of the spectrum.

The rules: what to try vs what never to touch

This is the article’s original element — the differentiator — so read it twice:

Green light (safe ground)	Red light (never)
Your own software and compiled projects	Software you do not own and have no right to inspect
Open-source binaries, CTF challenges, authorized test targets	A competitor’s app, cracked software, malware “for fun”
Static analysis first, runtime capture in an isolated VM	Runtime capture on your main machine with personal files nearby
Decompile → explain → recreate logic yourself	Decompile → lift code verbatim into your product

Good to know: runtime capture runs with your own user’s permissions — REA is not a sandbox (dev.to, 9 October 2026). If you cannot send decompiled code to a cloud model provider, run it inside an isolated VM with a local model, or stick to static analysis. Reverse-engineering someone else’s software can violate licences and laws; the safe ground is your own software, open-source binaries, CTF challenges, or targets you are authorized to test (dev.to, 9 October 2026). For the broader containment playbook — isolated environments, approval checkpoints, audit trails — see our AI agent safety checks guide. And the week this launched, an AI hacking tool was blamed for a breach at seven South Korean banks (Startup Fortune) — powerful tooling plus weak judgment is exactly the failure mode these rules exist for.

REA vs the earlier MCP bridges

Project	Stars (12 Oct 2026)	Scope
morluto/rea	~93,000 in a week	Many engines and formats, installer, CLI, evidence output
mrexodia/ida-pro-mcp	~12,600	IDA Pro only
LaurieWired/GhidraMCP	~11,000	Ghidra only

Star counts move daily — the point is the shape of the trade-off, not the scoreboard (agentbreaking.com read counts from the GitHub API on 12 October 2026). Single-engine bridges are older and more settled; REA is broader, faster-moving (v4.1, v5.0 and v6.0 shipped within three days, per dev.to) and demands the extra setup review.

Our take

The reason REA exploded is not that it clones software — it explicitly does not. It is the first popular packaging of the investigation model: a coding agent that cites function names, shows confidence levels and admits what it cannot see. That is genuinely useful for debugging legacy apps and learning from open-source code. But anything that decompiles software on your laptop and sends readings to a cloud model is a dual-use tool. Set it up in a VM, approve every diff it proposes, keep it pointed at code you own, and it is one of the best developer MCPs of the year. Skip those steps and you are running somebody else’s tooling with your permissions.

Frequently asked questions

Is REA free to use?

Yes — it is open source under the MIT licence, and setup runs from npx rea-agents setup (ai-tldr.dev, 6 October 2026). You still pay whatever your AI model provider charges for the tokens the agent uses.

Do I need Hopper, Ghidra or IDA installed?

For native-binary analysis, yes — REA talks to your existing installations (or can install Hopper during setup with your approval). Static JavaScript analysis needs no native backend (singulism.com, 10 October 2026).

Can REA recover original source code?

No — it explicitly does not claim to recover original source or clone whole applications (Startup Fortune, 9 October 2026). It produces readable decompilation, evidence bundles and explanations with stated unknowns.

Which agents work with REA?

Claude Code, Codex, Cursor, Gemini CLI and Grok Build are supported, with manual registration for other MCP-capable agents (singulism.com, 10 October 2026).

Is reverse engineering with REA legal?

It depends on what you analyze and where you live. Your own software, open-source binaries, CTF challenges and authorized test targets are the safe ground; pointing it at software you do not own can violate licences and laws. We are not lawyers — check your licence terms and local rules first.

Sources and methodology

Claims rest on five secondary sources checked 12 October 2026: agentbreaking.com’s 12 October star-count table (GitHub API read 12 Oct), Startup Fortune’s 9 October launch write-up, ai-tldr.dev’s 6 October v4.1 release notes, singulism.com’s 10 October setup walkthrough, and two dev.to analyses (9 October and 7 October). Star counts are moving targets and are presented as reported-at-date, not exact. Setup steps follow the project’s official guide as relayed by singulism.com and are desk-validated, not hands-on tested. Links open in a new tab.

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!