Matthew Berman promotes REA, which lets coding agents inspect apps without source code

Why it matters

REA pushes coding agents beyond writing from available source code and into examining shipped software. That expands what agents can help build, while making authorization and the limits of reconstruction central questions.

On October 8th, Matthew Berman (@MatthewBerman) pointed his followers to REA, an open-source tool that lets AI coding agents examine software without access to its source code. Berman called the moment "Software is done"; REA's own description is narrower: agents can inspect how a feature works and help build an adapted version, but the tool does not claim to recover original source code or automatically clone an app. REA's project page describes the workflow as "decompile, understand, recreate."

Berman is a promoter and user of REA, not its creator. In the six-post thread, he said he was using the tool and linked its repository. Berman's audience in AI brought visibility to the project, which is maintained under the GitHub pseudonym morluto. Berman has built a technology company before: he founded Sonar Technologies, a business messaging company, and wrote when Marchex acquired Sonar in 2019. He now publishes AI coverage and guides through Forward Future, where he said he may publish a full guide to REA.

The project turns specialist reverse-engineering steps into tools that a coding agent can call. REA's repository documents native binary analysis, decompilation and disassembly, string and symbol searches, call-relationship tracing, and analysis for JavaScript and Electron applications. It also offers a command-line interface and an MCP server, a way for compatible AI agents to use external tools. The workflow is designed to gather evidence from the shipped application, explain what that evidence shows, and guide a separate implementation. This gives an agent evidence to work from when analyzing an unfamiliar feature, instead of relying only on a model's guess.

REA is distributed under the MIT license, and its project documentation says analysis runs locally rather than uploading the application to a hosted analysis service. Deep native analysis can use Hopper or a user-provided Ghidra setup. The project describes its output as evidence for understanding behavior; it does not promise a faithful source-code recovery. An agent may help reproduce a feature, but REA does not establish that the recreation is complete, identical, or authorized.

Berman also offered a caution. In a reply, he compared the moment to the early "wild west of Napster" and predicted lawsuits. Tools that make technical inspection easier can also make it easier to study a competitor's implementation. Whether any particular inspection or recreation is permitted depends on the software and circumstances; REA's technical capabilities do not answer that question.

Generative coding tools are good at producing new code from prompts and existing repositories. REA targets cases where the code is unavailable and the behavior must first be inferred from an application or binary. REA's pitch is that agents can move from examining an existing product to building a comparable feature in the same development workflow, with evidence attached to the investigation rather than relying only on a model's guess.

Berman's "software is done" line is promotion, not a result demonstrated by the thread. REA is an investigation and reconstruction toolkit, and the project explicitly draws a line between that process and cloning. The project bets that coding agents can take on more of the work traditionally done by reverse engineers. Users' choices about what to do with knowledge of someone else's software remain unresolved.