fullauto.online

Tools

REA breaks the binary moat: 411,473 stars, 538,000 downloads a week, and what still protects your code

Reverse Engineer Anything was first committed on 14 April 2026. By 30 October it held 411,473 GitHub stars, and its npm package was downloaded 538,483 times in the previous week. Decompilers are not new — the NSA gave Ghidra away in 2019 — but nobody had put one behind a one-line install for coding agents. Here is what the tool does, what its own disclaimer rules out, what Thales and Quarkslab say actually stops it, and where UK, EU and US law draw the line.

Published
30 Oct 2026
Reading
10 min
Class
tools

half-life 90dfrom 30 Oct 2026

The GitHub API, queried on 30 October 2026, returns 411,473 stars and 102,975 forks for morluto/rea — a repository whose first commit is dated 14 April 2026 and whose last push landed the same afternoon (repo metadata read 30 October 2026). The npm package behind its installer, rea-agents, was downloaded 538,483 times between 23 and 29 October, and 1,172,649 times in the month to 29 October (npm registry API, read 30 October 2026). Whatever else is true about the star count, people are installing this.

REA — Reverse Engineer Anything — is a Model Context Protocol server and a command-line tool that hands your existing coding agent the standard reverse-engineering stack: Ghidra, Hopper, IDA, Rizin and GDB for native binaries, JADX for Android apps, Binwalk for firmware, TShark for packet captures, and a static-analysis path for JavaScript, Electron and .NET (project README, read 30 October 2026). Setup supports Claude Code, Codex, Cursor, Gemini CLI and Grok Build, plus anything that can register a local MCP server. Analysis runs on your machine, and findings come back with the evidence attached.

Its pitch is one line: “See a feature you like. Understand how it works, down to the binary level.” The uncomfortable part is that this describes something developers have wanted for thirty years and rarely got without a specialist and a fortnight. Here is where the project actually stands six months in.

How big it really is — and what the numbers do not tell you

The growth curve is the story everyone quoted, so it is worth putting the figures in one table with their sources attached. Star counts come from different collectors, so treat them as a shape rather than a scoreboard:

Date (2026)GitHub starsSource
8 October~20,000Mavday chart, from README history and GitHub API (21 Oct)
10 October66,886 (+25,784 that day)GitHub trending data via ComputeLeap/dev.to
12 October~110,000Mavday chart (21 Oct)
17 Octoberover 200,000InfoQ, 17 Oct
18 October319,213GitHub API, reported same day
21 October355,600Mavday, GitHub API 21 Oct
30 October411,473GitHub API, read for this article

Source: table compiled from the citations in each row; the 30 October figure is a direct read of the GitHub API.

Downloads are the better signal, because nobody has a compelling reason to fake a package mirror. The rea-agents package pulled 415,576 downloads in the week to 18 October (Mavday) and 538,483 in the week to 29 October (npm registry API, read 30 October 2026) — still climbing three weeks after the hype peak. The release log explains some of that: 39 published releases, with 4.0.1 on 5 October, 4.1.0 on 6 October, 5.0.0 on 7 October and 6.0.0 on 8 October — three majors in four days — and 6.8.0 most recently on 25 October (GitHub releases API, read 30 October 2026).

Two honest caveats. First, stars are noisy: a Hacker News commenter tracked what they called “an unusually high fork acceleration” back on 10 October, and Mavday, which looked, found no evidence of anything improper — 102,975 forks is simply what happens when a repo trends globally for a fortnight (Mavday, read 30 October 2026). Second, the project itself warns that “a similar repository name or copied README does not establish that a download is official” — lookalike repositories already exist, and the README names only three official sources: the GitHub repo, rea.tools, and the rea-agents npm package.

What it demonstrably does

The strongest evidence for what REA can do comes from the project's own showcases, because they publish the checking, not just the claim. The DX-Ball case study rebuilds a 1996 Windows game in maintainable C: 55 maintained functions, 45,380 comparisons against the original code, and 33 functions whose recompiled bytes match the original exactly, at a project checkpoint of 7 October 2026 (rea.tools showcase). The sound-panning function alone was verified across 3,205 original x86 test cases and all 63 compiled function bytes (InfoQ, 17 October 2026).

InfoQ's write-up also covers the Notion Electron clipboard investigation and a 16-bit DOS game reconstruction, and makes the point that matters about the workflow: the agent asks REA for a function, gets pseudocode, instructions, callers and raw byte reads back, and follows the evidence rather than improvising. That is why the output is checkable — a decompiler's first pseudocode view is often wrong, and the DX-Ball write-up shows the agent catching exactly that, a conversion call hiding the real input, and going back for the instruction view.

For the harness-minded, this is the pattern from harness engineering applied to an unfamiliar domain: the tool does not think, it returns evidence, and the agent's job is to reason over it. The MCP surface itself is worth understanding before you install anything — see MCP in practice and the four MCP attack surfaces.

What it does not do — read this before you install it

  • It does not recover source code. Compilation throws away names, comments and structure, so output is a reconstruction — and REA says so itself: the project “does not claim to recover the original source code” (InfoQ, 17 October 2026).
  • It grants you no rights. The disclaimer, last updated on 12 October 2026, limits the project to lawful research, education, interoperability and software preservation; it rules out “copying proprietary software or reproducing its protected code and assets” and states plainly that “obtaining source-like output or an analysis report does not grant rights to the original program”.
  • It supplies no tool licences either. Native analysis needs an existing Hopper, Ghidra or IDA installation, and the README is explicit that commercial tools must be legally obtained: cracked copies, stolen keys and licence bypasses are outside what the project authorises (README, read 30 October 2026).
  • Your model may refuse. Anthropic's usage policy bans infringing third-party intellectual property and finding or exploiting vulnerabilities without authorisation; the revision effective 12 November 2026 adds that it does “not prohibit security research, testing, or tool development conducted on systems you own or operate”. Developers report mixed results in practice, and REA's own disclaimer rules out evading a provider's usage policy (Mavday, read 30 October 2026).
  • Analysis runs locally, but the model still sees the findings. The disassembler never leaves your machine; the excerpts you ask it about go to whoever provides your model, subject to their data policy.

What actually stops it

Two independent write-ups are worth more than the star count, because both tried to break agent-driven analysis and reported what happened.

Thales launched Sentinel Envelope Plus on 1 October 2026, a paid add-on that transforms and recompiles selected parts of an application without source changes. In its controlled test, an AI agent found 8 of 10 vulnerabilities in an unprotected application; against the same application protected, it found none — after 970 times the token consumption, and the analysis was abandoned after nearly seven hours (Thales news release, 1 October 2026). Read that as a vendor-run test on one application: it measures effort, not impossibility.

Quarkslab ran the longer experiment. It spent several weeks handing sandboxed Claude Code sessions — full auto, permissions disabled, the strongest model available at the time — progressively hardened AArch64 binaries with one prompt: recover the hidden strings (Defeating AI-Assisted Reverse Engineering, 20 August 2026). Nothing handed over was deobfuscated inside an 80-minute budget, against an average of about 45 minutes for unprotected code. The failure mode it flagged matters more than the score: a stuck human says they are stuck, while “an agent evades impossibility by all means, and might produce a confident, yet hallucinated, report”.

Both reach the same conclusion from opposite directions. Protection multiplies attacker cost rather than blocking them; what AI changes is throughput, since agents run in parallel, overnight, with whatever tools they can reach. The defences that work are the old ones: bind secrets to the environment, make analysis require execution, make execution require the real thing.

Being one command away from a disassembler changes the economics of enforcement, not the statute. The distinctions that matter, as summarised by Mavday (read 30 October 2026):

  • UK: section 50B of the Copyright, Designs and Patents Act 1988 lets a lawful user decompile where it is necessary to build an independent program that can operate with the software, and section 296A makes a contract term purporting to forbid that void. What you cannot do is produce something “substantially similar in its expression”.
  • EU: Directive 2009/24/EC mirrors that, and the Court of Justice held in SAS Institute v World Programming (2012) that a program's functionality is not protected by copyright — copying what it does is different from copying how it is written.
  • US: Sega v Accolade (1992) treated disassembly as fair use where it is the only route to functional elements and there is a legitimate reason; Bowers v Baystate (2003) enforced a shrink-wrap ban on reverse engineering as a contract; DMCA section 1201 prohibits bypassing access controls with a narrow interoperability exemption.

Your licence terms and your jurisdiction both matter, and nothing here is legal advice. The practical rule from that split: studying a program you are entitled to examine is on far firmer ground than shipping a rebuild of it.

The clone wave is a separate story

REA's arrival coincided with an AI-built clone wave — open-source rewrites of Photoshop-class applications, including PhotoCraft, which describes itself as a clean-room Rust reimplementation built “from public specs and observed behaviour only”. Mavday found no evidence that PhotoCraft used REA, Ars Technica reported on 7 October 2026 that its developer used Claude Opus 5.5 and expected feature parity within a month while his own README calls the project early alpha, and Adobe has made no public response. Trade dress is a separate risk from copyright, and a claimed parity timeline from a maker is not a shipping product.

The distinction is worth keeping clean in either direction: REA is a measurement instrument that its own disclaimer limits to lawful uses, and the clones are the thing that will land in court first — if any of it does. For the agent-security picture more broadly, this sits alongside the LSCBench finding that agents write working-but-vulnerable code by default.

What is not settled

  • The star count will never be audited. Nobody outside GitHub can verify 411,473 stars, and the fork-acquisition rate drew comment on day one. Downloads are the verifiable figure: 538,483 in the week to 29 October 2026, and that number is public.
  • No defence result has been independently replicated. Thales published a vendor-run test (1 October 2026); Quarkslab published one lab's multi-week experiment on hardened AArch64 binaries (20 August 2026). As of 30 October 2026 no third party has reproduced either with REA in the loop.
  • No reported case yet. There is no UK or US decision involving an AI-assisted rebuild of a commercial program; the first one will say far more than any star count.
  • Provider policy is still moving. Anthropic's revised terms take effect on 12 November 2026; other providers have not published an equivalent clarification. A workflow that works this week may get a refusal next month.

What to do about it

  • Assume anything you ship can be read. API keys, licence checks and pricing logic inside desktop, mobile and edge apps were never hidden — they were just expensive to find. If any of them still live in a binary you distribute, move them server-side this month.
  • Try it on your own binaries first. The legitimate uses are real and they are the ones the disclaimer lists: a lost-source legacy executable, a closed file format you must read, a dependency you need to audit, software worth preserving.
  • Install from the three official sources only, and read what npx rea-agents setup writes before approving it — it registers an MCP server and installs workflow instructions into your agent's configuration, with backups of what it replaces.
  • Check the licence and the model policy before pointing it at someone else's software. Contract terms restricting interoperability are void in the UK and EU; licence bans can be enforced in the US; either way the analysis still has to stay inside what your provider's terms allow.
  • If you ship on-prem, embedded or edge software, budget for protection as a cost multiplier, not a barrier — and do not accept an agent's own report as proof that your protection worked. Quarkslab's hallucinated-success finding is the exact failure you would be relying on.
  • Track downloads, not stars. The npm figure is public, updated weekly, and cannot be inflated by a trending page.

Sources