REA MCP: reverse engineering tools for your coding agent
REA gives Claude Code, Cursor, Codex and other MCP clients a large set of reverse engineering tools for binaries, JavaScript apps and websites. Setup, the right MCP command, and what I saw testing version 6.9.0.
REA, short for Reverse Engineer Anything, is an open source CLI and MCP server from morluto that lets an agent look inside software it does not have the source for. You point your agent at an app, a binary or a website and it can read headers, imports and strings, rebuild the structure of a JavaScript or Electron app, decompile native code through an engine you already have, and report what it found along with the evidence and the limits of each answer.
It is the fastest climber on appsgit's weekly list of trending MCP servers, with about 17,700 GitHub stars gained between November 9 and November 16, and the repo showed a little over 460,000 stars when I checked today. It is MIT licensed and published on npm as rea-agents. I tested version 6.9.0, the latest release, on Linux before writing this.
What it covers
The catalog is wide. In 6.9.0 the MCP server listed 255 tools, grouped into families you can switch on or off. Some of what it handles:
Native binaries: ELF, PE and Mach-O headers, hardening flags, exports, imports, Go and Rust build metadata, plus decompiling through Hopper, Ghidra, IDA or Kuna if you have one installed.
JavaScript and Electron apps: a static graph of modules, entrypoints, network calls, IPC and preload scripts, built without running the code.
Websites: page bundles, network captures, screenshots and WebMCP tool discovery through a browser.
Android: APK inspection, JADX and apktool helpers, and adb device tools.
Runtime work: Frida sessions, GDB and Rizin debugging, process captures.
Static JavaScript analysis needs nothing extra. Native decompiling needs one of the engines above, and setup can offer to install Hopper with your approval.
Install
REA needs Node.js 22.23.4 or later on the 22 line, 24.22.1 or later on the 24 line, or 26.12.1 and up. The quickest route is the setup command from the README, which detects your agents, registers the MCP server and installs REA's workflow instructions, with backups of the config files it changes:
shterminal
1npx rea-agents setup
ShellUTF-8Ln 1, Col 21
The README also warns that copied repos with similar names exist, so only install from the morluto GitHub repo, rea.tools or the rea-agents package on npm.
Add it to Cursor by hand
The package ships this config. Note the mcp argument at the end. Without it, the command just prints the CLI help and exits, so your client will see a server that never starts.
1claude mcp add --scope user rea -- npx -y rea-agents@6.9.0 mcp
ShellUTF-8Ln 1, Col 63
Setup pins the exact version on purpose. When a new release comes out, run npx rea-agents@latest setup again to move the registration forward.
Keep the tool list small
With 255 tools, the size of the tool list matters. In 6.9.0 complete output schemas are advertised by default, and the full tools list response I got was about 5 MB of JSON. Two environment variables help:
REA_MCP_OUTPUT_SCHEMA_PROFILE set to off dropped the response to about 375 KB in my test, and compact brought it to about 690 KB. REA still validates its results either way.
REA_MCP_TOOL_FAMILIES limits which families are listed. Setting it to application and javascript-runtime gave me 37 tools instead of 255. An unknown family name stops the server at startup and prints the valid names.
The README on the main branch already says output schemas are left out by default, so this default may change in the next release. Check the README that ships with the version you install.
What I saw when I tested it
I installed rea-agents 6.9.0 from npm on a Debian machine with Node 24.22.1, started it with rea mcp and talked to it over stdio. I have no Hopper, Ghidra or IDA on that machine, so I stuck to the parts that need no engine.
The server started in under a second and reported itself as rea 6.9.0 with 255 tools.
rea doctor said healthy false, because no native analysis engine was installed. It named each missing engine and the environment variable or setup step that fixes it.
I wrote a 15 line Node app with two fetch calls to a made up API and an http server that answers two paths. analyze_javascript_application found the package entrypoint, the http import and both fetch calls, shown as API plus the login path and API plus the orders path with partial resolution, since the base URL lives in a constant.
It did not list the two paths my http handler checks as routes. The route inventory came back empty, and the output says plainly that missing facts do not prove a route is absent.
The full result for that tiny app was about 500 KB of JSON, mostly evidence ids, hashes and limitation notes. Passing detail set to summary returned a much shorter overview. Ask for the summary first.
inspect_binary_hardening on the system ls binary reported PIE, NX, full RELRO, stack protector and FORTIFY signs as present, CET as absent and the file as stripped. Each row gave the reason and the file offset it read.
Every result was tagged as untrusted content from the analysis target, which is a sensible hint to the agent not to follow instructions hidden inside the thing it is inspecting.
Things to keep in mind
Releases come out often. Pin a version, as setup does, and update on purpose.
The answers are careful about what they do not know. That makes them trustworthy, but also long. Summary views and smaller tool families save a lot of context.
Some tools are not read only. Debugger sessions, Frida, adb installs and pushes act on real processes and devices. Only connect REA to machines and devices you are allowed to inspect.
Reverse engineering other people's software can run into license terms and local law. The project ships a disclaimer in the package, and it is worth reading.
Find it on PlugMCP
REA is listed on PlugMCP with links to the repo, the npm package and the website. If you want your own agent to search the directory for servers like this one, the PlugMCP directory MCP needs no key.
ffmpeg-skill gives Claude Code, Cursor, Codex and other MCP clients 42 FFmpeg tools that cut, reframe, caption and check video on your own machine. The right MCP command and what I saw testing version 3.10.1.