Point an Agent at an App With No Source. It Will Tell You How It Works, and What It Could Not Prove.

On November 7 a TypeScript repository topped GitHub's daily star chart with about 3,400 stars in a day. It ships no weights and trains nothing. It is a tool that lets an agent take software apart.
The project is called rea, for Reverse Engineer Anything, and it ships as an MCP server. You install it, your agent gets a set of tools, and you ask questions like: how does search work in this app, show me the evidence, then build something similar for my project.
What comes back is the reason this is worth writing about. Every finding arrives with the evidence behind it and the limits of that evidence. Not a confident paragraph. A claim, the strings or pseudocode or headers it came from, and a note on what the tool could not see.
Breadth is the first surprise
The target list is long, and each entry implies a different kind of analysis.
Native binaries get pseudocode, assembly, strings, symbols and cross-references, through Hopper, Ghidra, IDA or Kuna. Go binaries get the compiler version, modules and build configuration, with no Go toolchain required. Rust binaries get crate information through cargo-auditable. GPU code objects get CUDA and HIP targets, kernel parameters and PTX text. Model files get GGUF and safetensors metadata. Unity games get IL2CPP metadata across versions 24, 27, 29, 31 and 39. Godot games get PCK archives unpacked, formats 2, 3 and 4.
Then the boring formats that actually matter: PE certificates and import tables, ELF notes and hardening flags including PIE, NX, RELRO and CFG, Apple IPS crash logs, Linux kernel module metadata, Erlang BEAM files, WebAssembly, EVM bytecode, SQLite databases, PyInstaller executables, Linux core dumps.
And the application layer: JavaScript and Electron bundles with modules, imports, source maps, routes and IPC channels. Websites, with page structure, scripts, network observations and screenshots. .NET assemblies with metadata, CIL instructions and native dependencies. Android APKs through JADX. Firmware through Binwalk and Unblob. Packet captures through TShark or HAR files. Live process behaviour through a native PTY.
Almost none of this is new capability in isolation; every one of those formats has a tool. What is new is that an agent can call them in sequence, keep the results in one context, and answer a question that spans several of them.
The install is one command, and it edits your config
Setup is `npx rea-agents setup`. It registers the MCP server with your agent, installs a set of workflow instructions, and backs up your existing configuration before touching it. Then you restart the agent.
Supported clients listed by the project include Claude Code, Codex, Cursor, Gemini CLI and Grok Build, with manual registration for anything else that speaks local MCP. There is also a global install, `npm install --global rea-agents`, which gives you a `rea` command for direct use.
The CLI path is useful when you do not want an agent in the loop at all:
``` npx -y rea-agents@latest analyze-javascript-application /absolute/path/to/app --json ```
That returns the same structured findings without a model interpreting them, which is the right mode for anything you want to script or archive.
There is a third mode that most coverage will skip. `rea environment-build` takes a HAR capture of a web session and builds an agent environment from it, then `agent-env run` executes it. The stated purpose is reinforcement learning environments, and it requires AgentEnv and Docker. Capturing a real interaction and turning it into a repeatable task is the least glamorous and most reusable thing in the tool.
The catch is the decompiler licences
Deep native analysis depends on external engines: Hopper, Ghidra, IDA or Kuna. Ghidra is free. The others are commercial, and the README says plainly that the project does not provide licences, keys or activation, and that using cracked builds, stolen keys or unauthorised shared licences is prohibited.
That paragraph is doing real work. A tool whose value proposition is opening software that is not yours, shipping instructions for pirating the decompiler it depends on, would be a liability. Drawing the line at the dependency is the difference between a research tool and a piracy kit.
The rest of the caveats are the ordinary texture of software that touches the system: Node version floors (22.23.4, 24.22.1, or 26.12.1), a procps-compatible `ps` on Linux for process attribution, WABT 1.0.42 for WebAssembly, JADX plus a full JDK for Android, Chrome for websites. Hopper may pop a demo-or-activate dialog on first run, which will hang an automated workflow until someone clicks it in the same user session.
Windows support for Ghidra is labelled experimental. Runtime capture, the mode where the tool interacts with a running process, executes with your user privileges and has side effects. The updater keeps at most 4 MiB of output, so very large metadata responses get rejected after the command exits.
Local analysis, with one honest asterisk
The tool analyses on your machine and does not upload the application. The README adds the sentence that most vendors omit: the agent receives the tool results, and the model provider has its own data policy.
That is the correct framing. Local inference is not the same as local analysis when a cloud model is reading the output, and the distinction matters to anyone evaluating this for a client's codebase. If the output cannot leave the building, the model cannot leave the building either.
The security note at the end is also worth repeating, because it is unusual: similar repository names or a copied README do not mean a download is official. Get it from the listed GitHub organisation, the project's own site, or the npm package.
Who this is for, and who it is not
The legitimate uses the README names are research, education, interoperability, software preservation, and software you own or are authorised to analyse. Interoperability is the one with the widest practical reach. If a file format is undocumented and you need to read it, pointing an agent at the application that writes it and asking what the bytes mean is faster than reverse engineering by hand, and produces a written record of how the conclusion was reached.
The illegitimate use is obvious and the tool does not pretend otherwise. What it does is put the boundary in the documentation instead of the marketing.
For a developer, the interesting shift sits in the output format: claim plus evidence plus limit. That format is what makes an agent's answer reviewable, and reviewable is the only property that lets this kind of analysis enter a normal engineering workflow.
A confident answer about how an app works is worth very little. An answer that shows you the four strings it inferred from, and says the fifth inference is a guess, is worth a week of work.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.