REA Tools: A Website for Agent-Driven Reverse Engineering

By Indie Kings | October 29, 2026

Updated October 29, 2026: rea.tools now presents the REA reverse engineering tool as a product with a guided setup, worked case studies and a feature for turning recorded web sessions into reinforcement learning environments. Its GitHub repository reports 405,415 stars and 101,556 forks as of today, up from the 391,795 stars this site cited on October 26, and the project remains MIT licensed.

Disclosure: this site covered REA on October 26 in an overview of the MCP server. This article focuses on the rea.tools site, its onboarding and the newer environment building feature.

Image: a Hewlett-Packard 1653B logic analyzer, via Wikimedia Commons.

The setup is the product

The site's pitch is one sentence: reverse engineering, with your coding agent. The longer version tells you to ask your agent about an app, a game or a website, and REA gives it tools to inspect the code behind the answer. That framing avoids the usual mistake of agent tooling, which is to present a pile of capabilities and leave the user to invent a workflow. Here the workflow is the front page, and the installation is written as a single instruction you paste into an agent rather than a sequence of terminal commands.

What the command actually does is more than install a package. It adds an MCP connection, which is how the agent calls REA, and an investigation skill, which is a procedure for using those calls. Existing configuration is backed up, and the setup shows a plan before applying changes. A verification command reports a healthy status with a JSON flag, which makes it usable in automation as well as in a chat window. The prerequisites are narrow: Node 22.23.4 or later on the 22 line, 24.22.1 or later on 24, or 26.12.1 or later, and one of the supported clients, which include Codex, Cursor, Windsurf, VS Code, Claude Code, Claude Desktop and OpenCode.

The honest way to judge an agent tool is to watch it fail. Every step of this flow has a known failure point: Node version mismatches, MCP servers that start but expose no tools, skills that the client does not load, and paths with spaces on Windows. The site addresses those directly with troubleshooting sections, a doctor command and per client setup IDs, which is a better signal than a feature grid. It also means the product is being used by people who hit those problems, because nobody documents them otherwise.

Case studies as proof

Three showcases on the site do the work that a demo video usually fails to do. The first traces the six digit login code in Aegis, an Android authenticator, through its released APK and lets you move a demo clock to see why the code changes every 30 seconds. The second follows Notion's packaged JavaScript to the system clipboard and shows how a copy preserves rich block data. The third recovers how a ball's position becomes left and right sound in DX Ball, a 1996 Windows game, which is a compact example of reading assembly for a specific calculation.

Each case study follows the same structure: a question, the evidence, and a reconstruction. The Windows Calculator example on the home page is the clearest. It asks why 200 plus 10 percent gives 220, shows the branch instructions that distinguish plus from multiply, traces two helper calls, and lands on the rule that percent takes a percentage of the first number after addition. The page marks which part came from REA and which part came from the agent, which is a useful honesty device: the tool supplies instructions, decompiled code and call references, and the model supplies the explanation.

That division matters in practice. A reverse engineering tool does not need to be intelligent if it can hand the model the right evidence at the right time, and a model does not need to be an expert if the evidence is well chosen. The risk is that a confident model narrates a plausible story that the evidence does not support, and the site's answer is to make the evidence visible next to the claim. For anyone who has read a hallucinated explanation of a function, that layout is the most persuasive thing on the page.

Recording an app to train agents

The feature that goes beyond reverse engineering is the RL environment builder. The idea is that your agent records a web app's requests and responses, then uses them to construct a local environment where other agents can practice tasks without running the original application. The site's example asks for a test app to be recorded across listing, creating, renaming and deleting folders, with a requirement that the recording show the folder before deletion and confirm it is gone afterward.

That requirement is the interesting part, because it is a correctness condition for the environment rather than the recording. If the environment cannot verify that a folder was deleted, it cannot score an agent on whether the deletion worked, and an RL environment that cannot score is just a mock server. Building these environments is normally a specialist job involving traffic capture, state modeling and reward design, so a guided path from a recorded session to a runnable environment is a real reduction in effort, provided the generated environment is faithful enough to be useful.

Whether it is depends on cases the site does not show, such as apps with authentication, background jobs or nondeterministic responses. The examples are deliberately simple, and that is fair for a first release. The larger point is strategic: the same instrumentation that lets an agent understand an app also lets it imitate the app for training, and that puts REA in the market for agent evaluation as well as reverse engineering. It is a smart extension of the core capability, and it explains why the site leads with outcomes rather than with the MCP protocol.

The numbers around the project deserve a note. GitHub reports more stars for REA than most of the largest open source projects in the world, for a repository created in April 2026, and a fork count above one hundred thousand. Those are the figures the API returns, and they have grown by roughly 13,600 stars in three days. Growth of that shape is unusual enough that readers should treat the counters as a marketing signal rather than a measure of adoption, and judge the tool by the case studies and the code instead.

Project card

FAQ

What is REA?
A tool that connects a coding agent to reverse engineering utilities through MCP, so the agent can inspect binaries, apps and websites and explain how a feature works.

How do I install it?
Run npx rea-agents@latest setup, or paste the setup prompt into your agent. It shows a plan, backs up existing configuration and adds the MCP connection plus an investigation skill.

Which agents are supported?
Codex, Cursor, Windsurf, VS Code, Claude Code, Claude Desktop and OpenCode, among others listed on the site.

Is it free?
Yes. The repository is MIT licensed, and the setup is distributed through npm and the project's own site.

What is the RL environment feature?
It records a web app's requests and responses and builds a local environment where agents can practice tasks and be scored without running the original app.

Can I use it on software I do not own?
The tool is general purpose, and legality depends on what you inspect and where you are. Owning the software or having permission is the safe baseline.

Bottom Line

The MCP server was already a good idea; the site is what makes it a product. Guided setup, a doctor command, per client documentation and case studies that show evidence next to claims are the difference between a repository and something a working developer will try on a Monday. The RL environment builder is the most forward looking piece, because it turns the same instrumentation into training and evaluation infrastructure. The extraordinary star and fork counts should be read with suspicion, and the showcase quality is a better guide. If you already use a coding agent for unfamiliar code, this is a natural extension of that workflow.

Sources

• rea.tools home page, get started guide and showcase descriptions, accessed October 29, 2026
• GitHub API metadata for morluto/rea, stars, forks, open issues, creation date and MIT license, accessed October 29, 2026
• Indie Kings coverage of REA from October 26, 2026, used for the disclosure note and the earlier star figure

Related: Filee: A Drag and Drop Converter for 266 Formats With No Uploads | Microsoft Database: Teen Exposes 17 Trillion Rows via Missing JWT