REA: how AI agents inspect applications without source code

What if you need to understand an application's logic but do not have its original source code? REA connects AI agents to specialist analysis tools, giving them evidence to investigate what happens inside.

From clicking to investigating internal logic

An AI agent can explore an application through its user interface: supplying inputs, observing results and inferring rules. This helps explain a workflow, but observation alone may not reveal an algorithm, hidden conditions or communication between application components.

REA (Reverse Engineer Anything) extends this approach. It is an open project with an MCP server that connects agents to specialist reverse engineering tools. MCP (Model Context Protocol) provides a shared interface for calling tools and working with their results.

Alongside observing the screen, an agent can inspect application structure, decompiled code or captured runtime behavior. The project supports agents including Claude Code, Cursor, Codex and Gemini CLI.

What REA provides access to

Native binaries: tools such as Ghidra, IDA and Hopper provide assembly, pseudocode, strings and relationships between functions.
JavaScript and Electron: analysis covers modules, imports, source maps and IPC communication between application processes.
Web applications: browser tools inspect page structure, scripts and network activity. WebAssembly analysis can use WABT tools.
Android applications: JADX and Apktool help inspect APK files, manifests, resources and decompiled methods.
Recorded network traffic: captures can be analyzed with tools such as TShark or mitmproxy.

Capabilities depend on the target application, operating system and available tools. Static JavaScript analysis does not require a native decompiler; deeper inspection of native binaries does.

Why evidence matters

The main benefit I see is access to concrete evidence: a function's pseudocode, relationships between modules or a communication trace. The agent can use these findings to ask follow-up questions and explain the basis for its conclusions.

Decompiled pseudocode is not the original source code, and analysis may be incomplete. A new implementation therefore needs verification with real inputs, including edge cases.

Illustrative workflow: select one application feature, ask the agent to locate its implementation and explain the conditions governing its behavior. Then build your own version for your project and compare the results across test cases. This can help with interoperability or maintaining older software.

Getting started

With Node.js and npm installed, run:

npx rea-agents setup

The setup wizard offers configuration for supported agents, MCP server registration and matching instructions. Review and approve the proposed changes, then restart your agent. Additional requirements depend on the analysis type; see the project documentation for current instructions.

Give the agent a specific question: “Find out how search works in this application. Show the evidence for your conclusions and propose a similar feature for my project.” A clearly defined goal makes both investigation and verification easier.

REA analyzes targets locally, but tool results are passed to the AI agent and their subsequent handling depends on its provider. Analyze software you own or are authorized to examine.

Video: REA and Claude Code

My Czech video explores “How AI agents clone applications without source code! (REA + Claude Code)”.

Watch the video on YouTube (in Czech) →