Skip to content

REA wraps reverse engineering tools in an MCP server for agents

The MIT-licensed TypeScript project registers an MCP server so coding agents can inspect binaries, apps and websites locally, with evidence attached to each finding.

Draftship Editorial4 min read

A GitHub project that puts reverse engineering behind MCP

On 2026-11-20 the morluto/rea repository describes REA reverse engineering MCP, an MIT-licensed TypeScript project that connects coding agents to tools for inspecting software without its source code. The listing shows 470,280 stars and 114,881 forks on that date, 10,535 stars more than six days earlier. The stated job is to take a feature you have seen in another product, work out how it is implemented, and produce a version for your own project.

The repository carries its own warning about distribution: it says to get REA from morluto/rea, rea.tools or the rea-agents npm package, and that a similar repository name or a copied README does not make a download official.

How setup and the agent workflow fit together

With Node.js and npm installed, npx rea-agents setup registers an MCP server plus matching workflow instructions, backs up existing configuration and asks for approval before applying changes. The agent needs a restart afterwards. The source says setup supports Claude Code, Codex, Cursor, Gemini CLI, Grok Build and other agents, and that analysis runs locally. Results are meant to include the evidence behind a conclusion along with its limitations, so the agent can ask follow-up questions, explain behaviour or write and test an implementation.

The same tools are available from a terminal. The README's example inspects an extracted JavaScript or Electron directory or ASAR file and returns modules, imports, Electron boundaries and evidence as JSON:

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

The README also gives a prompt example: ask the agent to explain how search works in a notes app, show the evidence, and build something similar. That is the workflow the project is selling, and it depends on an agent that can run local commands.

What the tooling can look at

Native binaries return pseudocode, assembly, strings, symbols, calls and references, and require one of Hopper, Ghidra, IDA or Kuna; setup can install Hopper with approval. JavaScript and Electron analysis needs no native engine. Websites are inspected through a Chrome-family browser. Android APKs need headless JADX and a full JDK, and Android devices need caller-supplied adb.

A large group of targets relies on built-in static inspection with nothing extra installed: Go binaries, Rust binaries, model artifacts in GGUF or safetensors format, GPU code objects, IL2CPP metadata, Godot PCK archives, PE imports and certificates, ELF notes and dependencies, binary hardening flags, Apple IPS crash files, Linux kernel modules, Erlang BEAM files and NuGet packages. WASM needs WABT 1.0.42. .NET assemblies are static only. SQLite, EVM bytecode and recorded Linux crash analysis each carry platform caveats in the README.

Base requirements are Node.js 22.x (>=22.23.4), 24.x (>=24.22.1) or 26.12.1+, plus npm. On Linux, Ghidra, JADX, firmware tool launches and static JavaScript or Electron analysis need a procps-compatible system ps for process ownership checks.

Limits and open questions

The README lists dependencies in detail but says nothing about how accurate the analysis is or how often a conclusion is wrong. It does not address the legal side of reverse engineering. No release date or version number appears in the text, and cost is not stated. Whether the star count reflects real adoption is not something the page discusses. Flutter appears in the target table, but the entry is cut off in the supplied text, so its scope is unclear. The project does advise keeping installations current, since releases carry frequent bug fixes, and notes that Bun global updates need Bun 1.2.17 or later.

There are also smaller configuration details. MCP advertises compact input schemas by default, and output schemas can be advertised within a 4,096-byte budget by setting REA_MCP_OUTPUT_SCHEMA_PROFILE=compact. Server-side validation is described as unchanged either way.

What it changes for builders

The useful shift is procedural. Instead of reading a blog post about how a feature might work, a builder can have an agent inspect the artifact and report which file, function or binary section supports each claim. What the README describes is an explanation, evidence and a new implementation, not the original source code. The quality of the result therefore depends on the agent doing the follow-up work and on the external analysis engine for native targets.

What it means for building apps with AI

Tools like REA assume an agent running on your machine with access to local binaries and a terminal. Draftship works differently: you describe an app in plain language and it writes, builds and hosts the code, with a built-in database and user sign-in. Paid plans add a choice of AI model, source code access, custom domains and GitHub push. If you are working in that environment and want to know which models a project can use, the docs on the AI chat and choosing a model cover it. Nothing in the REA material says it integrates with Draftship.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.