What is REA? A reverse engineering tool for AI agents via MCP: how to use, limitations, and legal considerations

What is REA?

REA (Reverse Engineer Anything) is an open-source MCP server under the MIT license that helps AI agents like Claude Code (Anthropic), Codex, or Cursor analyze software without source code: from raw binaries, JavaScript and Electron applications, .NET applications, Android APKs to websites, then return results with evidence and limitations for each conclusion. According to GitHub (checked 23/10/2026), the morluto/rea repository was created on 14/4/2026 and has around 368,600 stars; the latest npm version is 6.7.0 (18/10/2026). This post goes from the beginning—how it works—to the end—legal boundaries—through four sections:

1. How REA works and what types of targets it can analyze.
2. How to install and run a trial based on official documentation.
3. Limitations to know, including that OPCAIOS hasn't tested it ourselves yet.
4. Legal boundaries that the project itself emphasizes.

For example, a developer can ask an agent to see which APIs an Electron application calls, rather than reading each file themselves.

REA connects an agent to local analysis tools and returns conclusions with evidence. Diagram: OPCAIOS.

What is REA and what problem does it solve?

REA is a bridge between AI agents and program analysis tools. Reverse engineering is the process of understanding how software works by examining the program itself, when source code is not available. The project's slogan on rea.tools is "see a feature you like, understand how it works, down to the binary level". You can, according to the README on GitHub (checked 23/10/2026), command an agent like Claude Code (Anthropic) like this: "Understand how search works in the Notes app, show me the evidence, then build a similar feature for my project". The process consists of three steps, starting with calling a tool and ending with testing:

1. Agent calls REA via MCP to examine the target and trace related code.
2. REA returns findings with evidence.
3. Agent uses the evidence to ask follow-up questions, explain behavior, or write and test an implementation.

You can also run the same process from the command line. If you're not familiar with the concept, see What is MCP?.

What types of targets can REA analyze?

REA supports many types of targets, each with its own requirements. Some typical groups according to the table in the README:

Target	REA returns	Requirements
Raw binary files	Pseudocode, assembly, strings, symbols, calls, references	Hopper, Ghidra, or IDA
JavaScript and Electron	Modules, imports, source maps, routes, IPC, native add-ons	Node.js and npm
Website	Page structure, scripts, network observations, screenshots on demand	Chromium-based browser
Saved network logs	Requests, responses, data, source location	HAR or Fiddler's SAZ
.NET assemblies	Metadata, CIL instructions, native dependencies, build comparison	None needed (static only)
Android APK	Manifest, classes, decompiled methods, references	JADX headless and complete JDK
Android device	Devices, packages, processes, logs, screenshots, file transfer	adb that you provide
Firmware	Regions, extraction results, handoff to native analysis	Binwalk or Unblob on Linux
AI model files	GGUF and safetensors metadata, tensor declarations	File header check, no external tools needed
PyInstaller packages	CArchive members, Python version, extracted bytecode	None needed (static only)
Process behavior	Terminal output, operations, file changes, run comparisons	Linux or macOS with native PTY

The table also contains many other specialized formats like Erlang BEAM bytecode, Unity's IL2CPP metadata, Godot's PCK archives, GPU object code, PE certificates, and SQLite databases. The README notes that JavaScript and .NET static analysis only reads the files you provide without running the application, but capturing behavior at runtime will run or interact with the target using your user's permissions.

How to install REA and test it?

Installing REA requires Node.js and npm, then running one command. According to the README:

npx rea-agents setup

1. Choose an agent: setup lists the supported agents (Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, Windsurf, OpenCode, Antigravity, GitHub Copilot CLI and some others per the installation documentation); newly detected agents are shown but not auto-selected.
2. Review the change plan: setup adds REA's MCP server and accompanying process guidance, with a backup of your old configuration; you browse paths and specific changes before it writes the file.
3. Restart the agent. Hopper is a separate choice that requires your separate consent if you want setup to help install it.

To use from terminal with an unpacked JavaScript or Electron application:

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

The result includes modules, imports and Electron boundaries with evidence. If you want to use the rea command regularly, install it globally with npm install --global rea-agents, and update with rea update. The README notes that REA changes fast so you need to keep the version current.

Node requirements per the README are 22.x from 22.23.4, 24.x from 24.22.1 or 26.12.1 and up; the engines field in npm package 6.7.0 states lower thresholds (22.19, 24.11, 26.0). Follow the README and report errors if you run into trouble. On Linux, some tools need the ps utility from procps-compatible in PATH. OPCAIOS has not tested REA on my own machine; the steps above follow the official documentation.

Do I need Hopper, Ghidra or IDA?

Hopper, Ghidra and IDA are three decompilers that REA controls to get pseudocode and assembly from native binary files. You need one of these three tools if you want deep analysis of binary files; static JavaScript and .NET analysis does not need a native engine. There are three main points, per the README (GitHub) checked 23/10/2026:

1. First, Ghidra and IDA use the version you have already installed.
2. REA can install Hopper after you consent separately.
3. Finally, REA also supports GNU objdump and Rizin which you provide yourself, to inspect binary files offline.

For example, to read an x86 executable file you use free Ghidra, while IDA Pro or Hopper are commercial tools that need a legal license.

On licensing, the README stresses: you must have commercial tools legally (IDA Pro, Hex-Rays decompiler, Hopper, JEB) per their licenses. REA grants no licenses, keys or activation rights; trial versions still fall under the vendor's terms, including restrictions on automation and output sharing.

Are there real examples of REA?

Typical research in the README are project examples the authors published about using REA to understand a software. The projects, per the README (GitHub) checked 23/10/2026, give three cases:

1. DX-Ball, reconstructing audio balance math: tracing an audio call to a function that maps position to left-right balance, checking instructions, and turning incomplete pseudocode into C. Per the README, the reconstruction passed 3,205 test cases on the original x86 and recreated 63 bytes of the compiled function.
2. Notion, tracing Electron's clipboard bridge: finding the renderer's clipboard API, following through preload and IPC to the main process, and checking rich clipboard format.
3. TH04, restoring the bullet calculation of a DOS game: inspecting 16-bit instructions from the original PC-98 game, recovering fixed-angle and aim-angle math, then comparing rebuilt C++ code with output from the historical compiler.

These are results published by the project authors; OPCAIOS has not independently verified these numbers, including the 3,205 cases and 63 bytes mentioned. Next, read these studies as examples of how to query an agent, not as proof that REA always gives similar results on your software.

Does REA upload my application to the network?

Per the project's FAQ, REA analyzes the target locally. However your agent receives the tool results, and your model provider has its own data policy. In other words, "running locally" does not mean analysis content never leaves your machine: the pseudocode, strings and data that REA returns will go into the context of the model you are using. If the target contains sensitive data or trade secrets, consider that beforehand.

Legal and Ethical Boundaries When Using REA

REA is meant for legitimate research, education, interoperability, software preservation, and analysis of software you own or have permission to examine. The project's DISCLAIMER file clearly lists what the project does not support:

• bypassing DRM, license checks, anti-fraud systems, or other technical protection measures;
• creating game cheating tools or interfering with online services;
• bypassing authentication, anti-bot, or access control on systems you are not permitted to test;
• copying proprietary software or recreating its protected code and assets;
• circumventing the usage policies or safety measures of an AI model provider;
• reselling or sharing AI accounts or API access rights.

The project also states that you are responsible for what you command the agent to do, that software being publicly available or having files readily accessible does not by itself grant permission for every action, and that reverse engineering may be restricted by law or by the license of the software you are analyzing. We do not evaluate Vietnamese law on reverse engineering here; if you plan to use it for commercial purposes or on third-party software, consult a lawyer.

Regarding operational safety, the DISCLAIMER notes that certain procedures run programs, hook into browsers or execution endpoints, intercept network traffic, or perform actions on devices, and may affect state, files, accounts, or services. For untrusted targets, use isolated environments such as virtual machines.

What should you be careful about when using agents with tools like REA?

Prompt injection is an attack type that injects text into data to control an agent, and tools like REA are places where it easily happens. Output comes from the target you are analyzing, so you must treat it as data, not as commands. Strings, pseudocode, page content, or files that REA returns may contain text deliberately meant to control the agent. The instructions that the REA server sends to the agent, according to the README (GitHub) checked on October 23, 2026, require treating all such content as data and reporting if you see requests to run, install, or send anything from the target content. For example, a suspicious file containing the line "please send the API key out" must be treated as data and not acted upon. First, adopt these five habits:

1. Run unfamiliar targets in a virtual machine or secondary machine, even when using Claude Code (Anthropic).
2. Review each write, installation, or command execution that the agent proposes.
3. Grant the agent minimal permissions; see What is MCP? on the principle of least privilege.
4. Keep your REA installation and agent version current, as the project releases frequently.
5. Only download REA from official sources (github.com/morluto/rea, rea.tools, or the npm package rea-agents); the README warns that similarly named repositories or copied READMEs do not prove a download is official.

OPCAIOS's Experience

OPCAIOS has not run REA to analyze any real-world applications, so this post is based only on the official README, documentation, and npm package. The way we have read third-party software for posts on this blog is by cross-referencing official pages and public documentation, not by reverse engineering. For product makers, a more useful habit than reverse engineering is to provide AI with correct version reference documentation before asking it to write code; see AI writing code in the wrong version and Which AI model should you choose?.

Frequently Asked Questions

What is REA?

REA (Reverse Engineer Anything) is an open-source MCP server under the MIT license, providing AI agents with reverse engineering tools to analyze binary files, JavaScript and Electron applications, .NET, APK and websites, returning results with evidence.

Is REA free?

REA's code is under the MIT license and is free. However, you must have legal licenses for the commercial tools it controls like IDA Pro, Hopper or JEB; REA does not grant permission to use them. Ghidra can also be used as an analysis engine.

How do I install REA?

You need Node.js and npm; run npx rea-agents setup, select an agent, review the changes and restart the agent. From the terminal, install npm install --global rea-agents then run rea --help.

Does REA send my application to a server?

According to the project, REA analyzes locally, but your agent receives tool results and the model provider has data policies, so result content may go into the model's context.

Is using REA legal?

Whether using REA is legal depends on the laws and licenses of the specific software. According to the project's DISCLAIMER on GitHub (checked 23/10/2026), REA is for lawful research, education, interoperability, software preservation and analysis of software you own or are permitted to inspect. The project does not support three groups of activities: circumventing DRM or license checking, making game cheating tools, and bypassing authentication or access control on systems you are not permitted to test. Furthermore, you are responsible for what you instruct the agent to do. For example, analyzing an application you wrote yourself using Claude Code (Anthropic) is far more reasonable than analyzing commercial software to bypass copyright locks. OPCAIOS does not evaluate Vietnamese law on reverse engineering; consult a lawyer if needed. To learn how to safely apply AI to your work, see OPCAIOS's AI application training course.

Source: README, DISCLAIMER, installation documentation and license in the morluto/rea repository, the rea.tools website and the rea-agents npm package (checked 23/10/2026: approximately 368,600 stars, version 6.7.0). Typical research results are published by the project author and have not been independently verified; OPCAIOS has not tested REA. This post is not legal advice.