GTA San Andreas AI
I Reverse-Engineered GTA San Andreas With AI
What REA Can, and Cannot, Do
SEO title: I Reverse-Engineered GTA With AI: REA Tested
Meta description: I tested REA with Codex and Claude Code on GTA San Andreas. Here is what worked, what failed, the token cost, and why my demo stays private.
Last updated: October 14, 2026.
Eight hundred million tokens. One free repository. One classic game.
That was my test with REA, Codex, and Claude Code. REA means “Reverse Engineer Anything.”
I used my own legal copy of GTA San Andreas. I ran everything on my own PC.
After one week, I had a first demo prototype. It runs only on my computer. I have not published or shared it.
This article explains what worked. It also explains what did not work.
What is reverse engineering?
Reverse engineering means studying a finished product to understand how it works.
Think about a clock. You can open it and study each wheel. You did not receive the maker's plans.
Software works in a similar way. You may have the finished program, but not its source code.
Source code is the human-written text used to build software. Developers write it before making the final program.
The final program may contain a binary. A binary is machine-ready code your computer can run.
People study binaries to understand old software. They also use this work for security and compatibility.
The process is not perfect. A finished program loses many helpful names and comments.
You may see what a program does. You may not know why its developer chose it.
That gap matters. Evidence and guesses must stay separate.
What is REA?
REA stands for “Reverse Engineer Anything.” It is free and open source.
Open source means people can inspect and improve its public code. REA uses the MIT license.
Its repository gives a simple description:
Reverse engineer anything with agents, from app behavior down to native binaries.
An agent is AI software that can use tools and complete steps. REA connects these agents with existing analysis tools.
It works as a local MCP server and a command-line tool.
MCP means Model Context Protocol. It is a standard connection between AI and outside tools.
A command-line tool runs through typed computer commands. It does not need a normal app screen.
REA does not replace every analysis tool. It connects several tools through one workflow.
For native programs, it can use Ghidra, IDA, or Hopper.
Ghidra is a program for studying compiled software. It turns machine code into a more readable view.
REA can also work with JADX for Android apps. It supports Binwalk and Unblob for firmware.
Firmware is software stored inside hardware devices. TShark helps inspect saved network traffic.
REA supports many targets. These include:
- Native binaries
- Electron and JavaScript apps
- .NET assemblies
- Websites
- Android APK files
- Firmware
- SQLite databases
- Saved network captures
The evidence system is one of REA's best ideas.
REA can mark a finding as observed, inferred, or unknown. The README also explains evidence and limits.
Observed means the tool directly found supporting information. Inferred means the evidence suggests an answer.
Unknown means the available evidence cannot support a clear answer.
This does not make every answer correct. It makes uncertainty easier to see.
REA passed 200,000 GitHub stars during October 2026. That growth brought more attention to reverse engineering.
The repository also includes game examples. Its README shows case studies for DX-Ball and TH04.
How REA works
The basic flow has three parts.
1. An AI agent receives your question
You ask Codex or Claude Code about one feature. A broad request often creates broad guesses.
A focused question gives the agent a clearer search path.
2. REA calls the right analysis tool
REA links the agent with tools like Ghidra. The tool studies the selected program.
REA then returns evidence, limits, and remaining unknowns.
3. The agent explains or rebuilds the behavior
The agent asks more questions. It may then create a new implementation for your project.
An implementation is working code that performs a chosen task.
This new version is not the original source code. It is a reconstruction based on available evidence.
That difference is important. Similar behavior does not prove identical internal code.
My GTA San Andreas test
I tested REA with my legally owned GTA San Andreas copy. The work stayed on my own PC.
I used two AI coding agents. They were Codex and Claude Code.
Codex and Claude Code can inspect projects and write code. REA gave them extra analysis tools.
My goal was not to publish the game. I wanted to test REA on a complex target.
I focused on:
Maps
Movement
Camera
original Sound and story
I rebuilt the demo using:
Unreal Engine
I searched for a public source confirming this platform. I found no reliable result.
I do not want to guess. Add the exact tool before publishing.
The test lasted one week. The agents studied evidence and tried possible implementations.
The most surprising finding was:
It can provide a code guess accuracy almost 77% from the original code
The clearest wrong guess was:
It waste lot of tokes over 800 Million Token
That mistake belongs in this story. It shows why human checks still matter.
The AI bill
Tokens are small pieces of text processed by an AI model. A word may use one or more tokens.
Before this project, my lifetime total was about 300 million tokens.
After the project, it was about 1.1 billion tokens. The difference was around 800 million tokens.
My stats card also shows a 68 million token peak. It shows a longest task of 43 minutes and 2 seconds.
Those two numbers are lifetime statistics. They are not only from this project.
The token number does not prove that every step was useful. Large usage can include failed searches and repeated work.
The agents may inspect one path, find nothing, and try another. Long tasks can repeat this cycle many times.
This matters for anyone trying REA. The repository is free, but AI use may not be free.
Your cost depends on your model and plan. It also depends on task size.
Do not begin with a whole game. Start with one small behavior.
Set a usage limit before starting. Check the total after each major task.
Without limits, a research loop can consume far more tokens than expected.
The prototype
The result was a first demo prototype. It was rebuilt from what the AI learned.
It runs on my PC. I have not shared the code, files, or demo.
The exact rebuilt feature is:
[ADD DETAIL: Describe the finished prototype without including protected game material.]
Keeping it private was a deliberate choice.
Take-Two has acted against similar projects before. Take-Two is GTA's publisher.
In 2021, it sent DMCA notices covering re3 and reVC. These projects rebuilt parts of GTA III and Vice City.
DMCA means Digital Millennium Copyright Act. It supports copyright removal requests in the United States.
Take-Two later sued developers connected with those projects.
My project is not the same case. Still, that history is important context.
Reverse-engineering rules differ by country and purpose. Ownership alone may not answer every legal question.
This article is not legal advice. Check the law where you live.
The honest part
REA is useful, but it is not magic.
It makes existing tools easier for AI agents to use. Many core tools existed before REA.
That leads to fair criticism. Some analysts see REA mainly as useful packaging.
I think both views can be true.
Packaging matters when it saves manual work. Yet the underlying analysis still needs checking.
REA can help an agent search and connect evidence. It cannot recover every original design choice.
AI-written behavior may look correct during a short test. It can still fail in unusual cases.
The reconstruction may also differ from the original. Matching the outside result does not prove identical code.
The process can burn many tokens. My own project used about 800 million.
Large targets also create more places for mistakes. One wrong inference can affect later work.
The labels help, but they do not remove this risk. Review every inferred claim yourself.
Only study software you own or may legally examine. REA's README says the same thing.
Commercial tools also need valid licenses. That includes paid tools like IDA Pro.
Who should use REA?
REA can help several groups.
Developers maintaining old software
You may need to understand a program without its source code. REA can organize the first investigation.
Security teams checking their own products
A security team may inspect its own released software. The goal can be finding exposed behavior.
Only test systems you own or may examine.
Founders studying one product feature
A founder may like one feature in another app. REA can help explain its visible behavior.
The safe goal is learning the idea. Build your own version with your own code and design.
Do not use REA as a shortcut for copying a full product.
Researchers and preservation teams
Old software may lack useful documents. Evidence-backed analysis can help explain how it worked.
How to try REA safely
Keep the first test small and lawful.
- Open the official REA repository.
- Read the current README and installation guide.
- Use only software you own or may examine.
- Choose one small feature, not a whole product.
- Run the official setup command:
npx rea-agents setup. - You may install the CLI with
npm install --global rea-agents. - Review every proposed setup change before approving it.
- Check every inferred finding against direct evidence.
- Keep a list of unknowns and failed checks.
- Stop when the legal or technical boundary becomes unclear.
The setup may change. Always check the official README first.
Do not upload private customer software to an AI provider without permission. Check your model provider's data policy.
FAQ
Is reverse engineering legal?
The answer depends on location, purpose, license terms, and method.
Owning software does not automatically allow every use. Get legal advice for serious or public projects.
Do I need to code?
REA can reduce manual work. It does not remove the need for technical judgment.
You should understand files, programs, and basic code. Native analysis needs deeper skills.
What does REA cost?
REA is free under the MIT license. Other costs can still apply.
AI models may charge for tokens. Tools like IDA Pro may require paid licenses.
Hardware time also has a cost. Check every provider before starting.
Can REA copy a whole game?
REA can analyze evidence and help reconstruct selected behavior. That does not mean perfect copying.
A complete game contains code, art, sound, writing, and many connected systems.
Technical ability also does not grant legal permission. Keep the scope small and lawful.
Is REA safe to run?
REA analyzes targets locally. Your AI provider still receives tool results.
Runtime analysis may run the selected program with your user permissions. Use a separate test environment when possible.
Read the official security and setup notes. Review every command before approving it.
Final take
REA helped Codex and Claude Code study a complex game. It also exposed the limits of AI reconstruction.
The first demo took one week. It also consumed about 800 million tokens.
That tradeoff is the real story.
REA can shorten parts of an investigation. It cannot replace evidence, testing, or judgment.
Start small. Check every inference. Keep your legal boundary clear.
Follow @salama_win on Instagram for no-fluff AI experiments.
Sources
Stay ahead of the curve
Join my private newsletter for exclusive insights, tools, and thoughts straight to your inbox. No spam, just value.