REA (Reverse Engineer Anything) is a free, MIT-licensed GitHub project that connects AI coding agents like Claude Code, Codex, and Cursor to the tools that take compiled software apart. An agent can use it to study how an app feature works without the source code, then report what it found and how sure it is.

Software you install usually arrives compiled, meaning the developers' readable code was converted into machine instructions that a computer can run and a person can't easily read. Taking it apart to learn how it works is called reverse engineering, and it has always taken specialist skill and a lot of time. REA hands much of that work to an AI agent.

The project lives at github.com/morluto/rea. Reports put it under 10,000 stars on Oct 6, 2026 and above 100,000 by Oct 12. GitHub's page showed about 269,000 stars and 67,000 forks on Oct 15. A star is GitHub's bookmark button and a fork is a personal copy of the project, so both measure attention rather than installs, and one analysis suspects some of the growth came from drive-by stars.

The attention comes from one question: if an agent can learn how a paid app works, can someone rebuild it and give it away? The evidence so far points to small, testable pieces of software, and the best-known open-source clones of Photoshop and Word were built a different way.

What does REA do?

REA gives an AI agent a set of tools to inspect an app, follow the evidence, and explain a feature, then helps it build something similar. The README describes three stages:

  1. Decompile. Open the app and recover readable code, strings, and names. Decompiling converts a compiled program back into code a person can read, but the result is an approximation called pseudocode, not the original source.
  2. Understand. Follow the code across the app until the agent can explain how a feature works.
  3. Recreate. Turn the findings into a feature in your own product.

The README is direct about the limit: REA "does not claim to recover original source code or automatically clone an application."

REA is an MCP server. MCP, the Model Context Protocol, is a standard that lets an AI agent plug into outside tools (the beginner's guide to MCPs covers it). The server sits between your agent and engines like Ghidra, a free open-source decompiler, and Hopper and IDA Pro, which are paid. Each finding comes back with its evidence and limitations, so you can see what the tool observed and what the agent guessed.

It can inspect:

  • Native programs for Mac, Linux, and Windows
  • Electron apps, which package a website into a desktop app (Notion is one)
  • .NET programs, Android apps, firmware, and websites opened in a Chrome-family browser
  • Running programs, through a process capture mode that runs a program and records what it does

A tester counted 170 tools in version 6.4.0, the release dated Oct 11. No AI company makes REA. The copyright belongs to the GitHub account morluto, and none of the coverage reviewed for this post names the person behind it.

How are people using REA?

The documented uses are small and testable. Each one is a single feature or function with a clear right answer.

  • DX-Ball, a 1996 Windows game. REA recovered how the game pans sound left or right based on the ball's position. The rebuilt function passed 3,205 test cases from the original program and reproduced all 63 bytes of the compiled function, according to the project's own case study. The full game has not been rebuilt, per IAMag.
  • Aegis, an Android authenticator app. An agent traced how the app produces its six-digit login code, which changes every 30 seconds. A reconstruction matched the reference across 133 test cases, according to KAD.
  • Notion's desktop app. REA followed the app's packaged JavaScript to the system clipboard to show how a copy keeps rich block data. That was a tracing exercise rather than a rebuild, and IAMag reports one step was done by hand.
  • Windows Calculator. An example on rea.tools explains why 200 + 10% gives 220 and rebuilds the percent button.
  • Loom's recorder. MindStudio describes creators rebuilding Loom's camera bubble, countdown, and pause and resume as a local recorder that saves to a folder instead of the cloud. The clones looked visually different from the original.
  • ArtCraft's open-source Photoshop and Word clones. These are the projects many people mean when they say AI is cloning paid software. ArtCraft is a team that publishes free, open-source Rust reimplementations of Adobe and Microsoft apps, such as PhotoCraft and WordCraft. They are tied to REA by timing and by theme, not by a documented link. PhotoCraft's README says it was "implemented from public specs and observed behaviour only," and neither README mentions REA. The next section covers what they are and how far along they are.
  • A TV and a media app. One user on X reported pointing it at an LG TV and turning the result into a Plex client. That is a single unverified report.

The five-step workflow people follow

MindStudio describes a method creators use with REA:

  1. Pick one specific feature, not a whole app.
  2. Identify the likely tech stack before investigating.
  3. Condense the raw analysis into a shorter summary and save it locally.
  4. Tag every finding as seen, inferred, or unknown.
  5. Rebuild the feature and test it against the original's behavior until it matches.

Step 4 exists because agents can state a guess with the same confidence as a verified fact, and the mistake shows up later as a bug that is hard to trace.

What a rebuild costs

Stork.ai ported the 1995 game Space Cadet Pinball to Rust with Opus 5.5 in about 2 hours 25 minutes. The port came to roughly 31,000 lines of Rust and 174 tests, and cost about $140 in API fees. The article contrasts this with REA and does not say REA was used, but it gives a rough price for a bounded program with clear pass or fail behavior.

Is REA behind the open-source Photoshop and Word clones?

The people who built the best-known clones say they used a different method. The clones come from a project called ArtCraft, started by a developer named Brandon, and they are written in Rust (a programming language). The ones I checked are dual-licensed under MIT or Apache-2.0, both permissive open-source licenses. ArtCraft is also the name of the team's separate AI image and video studio app, so the open-source clones are easiest to find under their own names. According to IAMag, 13 projects appeared between Sept 30 and Oct 9, including PhotoCraft (Photoshop), WordCraft (Word), and GridCraft (Excel).

ArtCraft says it worked from public file format specifications (Adobe's PSD spec, and OOXML for Word documents) and from watching how the real apps behave. WordCraft's rules forbid reading or disassembling Word. Brandon has said he was motivated by Adobe's cancellation fees, according to Stork.ai. PhotoCraft sits at about 63,000 stars, and its README calls it "early alpha" and "not yet a Photoshop replacement for daily professional work."

Where to look at them yourself:

The clones are unfinished, and their creators say so:

  • GridCraft implements 513 of Excel's 523 functions and is about 50% ready for production, by its creators' estimate.
  • WordCraft covers 88% of Word's ribbon commands and is about 60% of the way to replacing Word.
  • PhotoCraft is not yet viable for daily professional work, per IAMag. ArtCraft's roadmap puts Photoshop parity "well below 50%," according to Stork.ai.
  • Early Hacker News testers reported misaligned text selection, shortcuts that didn't work, and unresponsive sliders.

A viral version of the story says an agent was pointed at a local copy of Photoshop, clicked through every panel, and rebuilt it. MindStudio's Photoshop article tells it without naming a builder or linking a repository. The same article reports that a general coding agent with no reverse-engineering library produced a comparable clone of a simple Mac photo viewer, and that the REA version had some extra fidelity but took longer.

So the answer to "how are people using REA to open source software" is narrower than the headlines. The documented use falls into three groups: understanding and documenting one feature, rebuilding a function or a small app, and software preservation work like the DX-Ball and TH04 (a PC-98 game) projects. A full open-source replacement for a paid suite has not come out of REA.

Is it legal to rebuild software this way?

It depends on the method, the country, the terms you agreed to, and what you copy. The case law that exists predates AI agents, and none of it involved an AI rebuild, according to one legal analysis published Oct 14.

What courts have said so far:

  • Sega v. Accolade (1992) and Sony v. Connectix (2000): disassembling software to reach its unprotected parts was fair use.
  • Google v. Oracle (2021): copying about 11,500 lines of Java interface declarations was fair use (a legal allowance to reuse copyrighted material in limited ways), with the court assuming without deciding that they were protected.
  • SAS v. World Programming (EU, 2012): a program's functionality and file formats are not protected by copyright, and observing software to understand it is allowed. Copying the manual can still infringe.

What can still cause trouble:

  • Terms of service. Adobe's terms of use, in force since October 2025, bar recreating its products, according to the same analysis, which found no public Adobe action as of Oct 13.
  • Anti-circumvention rules, which can apply when a project bypasses DRM (copy protection) or other technical protections.
  • Trade secrets and confidentiality agreements.
  • Trademarks, trade dress (the look and feel of a product), and patents.
  • Copied assets such as icons, fonts, and sample files.

Many serious projects use a clean-room process to reduce the risk. In the classic version, one team reads the original and writes a specification, and a second team codes only from that specification, with the exchanges logged. Phoenix Technologies used it for an IBM-compatible BIOS in 1984. The AI version is looser, because one person often observes, specifies, and reviews, and the wall between them is a prompt and some permission settings. A 2026 rewrite of the chardet library shows the problem. Its maintainers rewrote it with Claude Code in five days and moved it from the LGPL license (which keeps reused code under similar terms) to the more permissive MIT. A contributor objected that the rewrite was not a clean room. The maintainers' published session transcripts show a helper agent read the old code on its own, yet similarity tools scored the new code as very different. The dispute is unresolved.

REA's own disclaimer prohibits circumventing DRM, license checks, and anti-cheat systems, and puts responsibility on the user. It also says that "public availability of software, possession of a file, or access to a debugging endpoint does not by itself establish permission for every operation."

This section is information, not legal advice. If you plan to publish something built from another company's product, talk to a lawyer first.

What can't REA do yet?

REA's static analysis (reading code without running it) is fast and its results are honest about gaps, but it misses things, and it cannot see code that runs on someone else's server.

HackerNoon tested version 6.4.0 on a real Electron app with a 100 MB package. The scan took about 52 seconds and parsed 421 JavaScript files with no failures. It reported zero preload scripts when the app has two, and it reported 744 limitations. A search for a channel name that appears in the source returned nothing, and the tool noted that "no match" is not evidence of absence. On a small test app, counts were nearly all correct, but it missed a sender check done inside a helper function.

Another test of version 6.4.0, reported by AgentConn, got all 30 file-and-line citations right and still missed items such as all 4 of one Electron app's nodeIntegration settings (a security switch). The warning that comes out of both: missing findings don't establish absence.

Cost is the other limit. The tool list alone ran about 1.95 MB in the HackerNoon test, which the author estimated at roughly half a million tokens (tokens are the chunks of text a model reads and bills by). One summary call took 60.5 seconds and returned about 1.35 MB. Watch your usage if an agent is calling it.

Verification is the bottleneck. A rebuild that passes its tests can still be wrong if a single constant is off, and a binary shows what code does but not why an edge case exists.

How do you install REA?

I'm not a developer or a programmer. From what I've been able to see, these are the basic steps, and if you ask me anything beyond them, I don't have more answers. The README and rea.tools cover the rest.

Setup runs from one command. Native binary analysis needs a decompiler installed as well.

  1. Check requirements: Node.js 22.19 or newer (or 24.11 or newer), and macOS 12+, Ubuntu 24.04+, Fedora 41+, or 64-bit Arch Linux.
  2. Run npx rea-agents setup. It shows the changes it plans to make, asks for approval, and backs up existing configuration.
  3. Restart your agent.
  4. Run npx -y rea-agents@latest doctor to check your environment without changing anything.
  5. For native programs, install a decompiler: Hopper (setup can install it on Mac and supported Linux, with your approval), Ghidra 12.1.4 with JDK 21, or IDA.

You're done when you can ask your agent to inspect a small app you own and it comes back with findings that cite evidence. Start there, on software you can check yourself.

Is REA free?

REA is free under the MIT license, which lets anyone use, change, and share the code. The engines have their own terms: Ghidra is free, Hopper has a demo with vendor-set limits, and the README says you must legally obtain commercial tools like IDA Pro and Hopper. You also pay for whatever AI agent runs it.

Does REA work on Windows or Mac?

The README lists Mac and Linux as supported hosts. It does not list Windows, and Ghidra operations on Windows are blocked until the project adds process-ownership and path-safety controls. The HackerNoon tester ran static analysis on Debian 13 even though rea doctor marked the host unsupported. AgentConn counts 8 releases between Oct 3 and Oct 11, so check the README before you rely on any of this.

Safety notes

  • REA is not a sandbox. A launched program runs with your permissions, so use a virtual machine for anything untrusted.
  • The model provider behind your agent receives the tool results, even though the analysis runs locally.
  • The README suggests a curl ... | bash installer. A SkillsLLM scan flagged that as low severity; read the script first, or use npm.
  • Install only from github.com/morluto/rea, rea.tools, or the rea-agents npm package. Lookalike repositories appeared within hours of the project trending, according to AgentConn.

What could REA mean for software companies?

Software that runs on the customer's device is the exposed part. Desktop apps, mobile apps, Electron apps, and the front end of a website all ship to the customer, and until now the cost of the skill needed to read them was a form of protection. AgentConn argues that this protection came from the price of the skill, and that REA lowers the price without changing the law. PlayCISO draws the conclusion for defenders: assume someone will read your shipped app.

Obfuscation, which scrambles code so it is harder to read, slows an agent down. PlayCISO treats it as a delay. ComputeLeap reports a vendor-run benchmark of one obfuscation product that raised the cost of analysis sharply, and the vendor itself says only that attacks become "harder and more expensive."

Practical steps for any company that ships software:

  • Run REA against your own app first and read what it finds.
  • Move business logic, license checks, and feature gates to your servers.
  • Treat any API key or token in a shipped client as already exposed.
  • Write a policy for who may run agent-driven reverse engineering and on what.

The business pressure shows up first in pricing. In mid-October, HSBC downgraded Adobe from Buy to Hold and cut its price target to $267 from $310, citing reports of ArtCraft's free clones, according to NAI500. The same report put Adobe's stock at $238.14, down nearly 32% for the year. A clone can matter well before it matches Photoshop. Computing speculates that 70% to 80% parity could satisfy many freelancers and small businesses, while the ArtCraft creators' own estimates run from well below 50% for Photoshop to about 60% for Word.

What could REA mean for SaaS platforms?

SaaS, software you rent through a browser and a subscription, is less exposed to REA than desktop software, because its core logic runs on the vendor's servers. REA can inspect a page's front end and watch network behavior in a browser, but it has no way to read code that never leaves the vendor's infrastructure. ComputeLeap argues that the logic that never leaves your infrastructure is what stays defensible. Data pipelines, models served through an API, and network effects fit that description.

The exposure is in what a customer can see: screens, workflows, file formats, and the shape of the API (the set of requests other programs use to talk to the service). KAD's point is that competitors, or a customer's own developer, can now reproduce the specific functions a customer uses without copying a whole product. For thin SaaS products, which are mostly a database with an interface on top, that matters. VibeOffice, a browser-based office suite, reported about $1,000 in model usage for its interface and core features, with much more spent on compatibility work (self-reported, per KAD).

The code is only part of what customers buy. Fortune's February 2026 reporting argued that many large companies won't build their own CRM or HR system, and that incumbents already own the data, access controls, and governance agents need. Building a clone is easier than keeping it correct, secure, and supported for years. Salesforce, ServiceNow, and Microsoft have already moved toward consumption and value-based pricing to answer the seat-license risk. The market reaction in early February wiped out roughly $2 trillion in software market value, per a Goldman Sachs estimate Fortune cited, and Computing reports that the sell-off eased by summer as investors saw some vendors benefiting from agents.

My read is that REA adds pressure in three places, and none of them ends SaaS:

  • Price. Seat-based pricing gets harder to defend when the value is mostly interface.
  • Build or buy. A team can rebuild the one feature it uses instead of renting the whole product.
  • Contracts. Expect more terms of service that ban reverse engineering, and more logic moved to servers.

What to watch next

These are the signals that will show whether REA's attention turns into real change:

  • Lawsuits and takedowns. As of Oct 13 there was no public action from Adobe. One analysis expects well-funded litigation once companies see core products threatened.
  • Changed terms. Watch for new restrictions in software licenses and terms of service.
  • A clone that reaches daily use. The creators' own parity estimates run from well below 50% to about 60% so far, and none has been independently verified.
  • Real usage. Stars measure attention. Watch npm downloads and how many projects cite REA in their own repositories.
  • Model refusals. One report says some frontier models refused reverse engineering requests. How AI providers handle this will shape who can use the tool.

Sources:
REA on GitHub (morluto/rea) · rea.tools · REA's disclaimer · ArtCraft · storytold on GitHub · PhotoCraft · WordCraft · I Tested REA on a Real Electron App (HackerNoon) · Reverse Engineer Anything: The GitHub Repo That Clones App Features With AI (MindStudio) · AI Reverse Engineering: REA and Office Clones (IAMag) · AI Reverse Engineering and Clean-Room Rewrites (Redreamality) · REA Hit 248K Stars (AgentConn) · Reverse Engineer Anything (REA) and What It Means for Defenders (PlayCISO) · AI Reverse Engineering: Is Software's Subscription Model at Risk? (KAD) · AI Cloned Your App Overnight. Now What? (DEV Community, ComputeLeap) · AI Cloned Photoshop. What's Left of the Software Moat? (MindStudio) · AI Agents, ArtCraft, and the Future of Software Cloning (Stork.ai) · Adobe's HSBC Hit, ArtCraft's Free Clone, and the Market's Mood Swing (NAI500) · This time, software really might be over (Computing) · AI agents aren't killing SaaS (Fortune) · REA Gives Claude Code and Codex an MCP Server (ccleaks)