Connect REA to your agent
Try it without installing, using npx, run the setup with --dry-run, and check only what you will use with doctor.
REA Training · hands-on track in 5 guides
REA connects Claude Code or Codex to reverse-engineering specialist tools such as Ghidra. In 5 hands-on steps you install it, point it at a program you are allowed to analyze, read the evidence behind each finding, and rebuild a function in your project, with your own code.
Guide in PT, EN, and ES. Videos in Portuguese. REA is a third-party project, MIT; INEMA made the guide, the tests, and the videos.

The problem
With the original program beside it, the agent has a right answer to compare against. Without a method, the problems below show up.
After the training
Try it without installing, using npx, run the setup with --dry-run, and check only what you will use with doctor.
Electron and JavaScript through REA's own analyzer, for free; native binaries through Ghidra, Hopper or IDA.
Each finding marked as seen, inferred or unknown, with the quoted excerpt and the limitations stated.
Summary first, then one module, one function, and each finding becomes a test against the original.
An open model running locally on Ollama or on a prepaid platform in the style of OpenRouter, knowing when the code leaves your machine.
What runs on the server can't be opened: the database, pricing rules, prompts and payments stay out of reach.
5 guides in order · from concept to real test
There is no course: the training is the INEMA guide, read in order, with the step-by-step, the method, the path without refusals and our real test. The videos go along with each part.
Guide
The bridge between your agent and Ghidra: it opens, understands, proves and rebuilds, as a CLI and as an MCP server, and what stays out of reach.
Guide
Prerequisites, the cost of each decompiler and the 4 steps: analysis without installing, setup with --dry-run, doctor with a scope, and the request to the agent.
Guide
One function, not the whole app; summary first; seen, inferred or unknown; and each finding tested against the original.
Guide
Open models without refusals, dsh and the local models tested, and two prepaid platforms for those who don't have a machine.
Real-world test
The Electron app that hung on 6.3.0 analyzed in 94.5 s on 6.4.0: 202 files read, 0 failures and 13,500 findings.
Who it's for
Included in the training
The guide, the test results and the script to run it isolated are in the INEMA repository. The official repository and the REA site come along, with the videos from the concept to the technical.
Why this training
We ran REA isolated, with a memory and time cap. On 6.3.0 an app hung; on 6.4.0 the same app was analyzed in full.
In a third-party test, the agent without REA finished first. For simple apps, sometimes it isn't even needed, and the track says so.
Only programs that are yours or that you have permission for, an isolated test folder, and reading the source of each finding.
A Windows walkthrough starting from Obsidian for those who don't program, and REA 6.4 under the hood for those who do.
Questions
No. You need an agent, Node.js and a program you are allowed to analyze. REA operates the decompiler for you.
1 → 2 → 3 and then 4, if your agent refuses. 5 shows what to expect from a real test. Start with the 6-minute video if you want the overview.
Node.js 22.19+, 24.11+ or 26+ (versions 23 and 25 are not supported) and Claude Code, Codex or another agent with MCP. For a native binary, Ghidra, which is free, or Hopper or IDA.
Yes. The track has a Windows walkthrough starting from Obsidian. If you hit an error, use WSL or a VPS.
Only with programs that are yours or that you have permission for. Usage licenses may prohibit reverse engineering. Studying for personal use is the safe ground; distributing or copying code, icons and texts is not.
Start with the guide. With Node installed, the first test runs with one npx, without Ghidra and without running the app.