---
title: "DeepSeek V4.1 Flash used REA to reverse-engineer a 1984 DOS game and rebuild it for Chrome, for $0.68"
url: "https://reamcp.dev/en/cases/deepseek-v4-1-flash-used-rea-to-reverse-engineer-a-1984-dos--229f34"
source: "https://www.reddit.com/r/DeepSeek/comments/1xr3g3i/deepseek_v41_flash_used_rea_to_reverseengineer_a"
source_type: "reddit"
author: "pzarevich"
author_handle: "pzarevich"
stars: 0
category: "App feature recreation"
categories: ["app-recreation"]
tags: ["binary-decompilation", "dos-game", "game-port", "ghidra", "html5-canvas"]
added: "2026-11-09T04:41:30.169Z"
updated: "2026-11-08T17:11:39.000Z"
locale: "en"
content_locale: "en"
---

# DeepSeek V4.1 Flash used REA to reverse-engineer a 1984 DOS game and rebuild it for Chrome, for $0.68

A developer describes using REA as an MCP server with Ghidra to decompile the 1984 Sopwith DOS binary, extract sprites, terrain and game tables, and rebuild it as an HTML/Canvas game, with repository links and methodology.

## Description

The Reddit post details a concrete REA workflow: DeepSeek V4.1 Flash used REA, an MCP server running Ghidra, to decompile a 43 KB DOS game binary, recover its game loop, physics, scoring and data tables, and reconstruct it as a playable HTML/Canvas port. The author includes links to the open-source agent harness and REA, describes extracting assets directly from the EXE, and compares the port against the original running in DOSBox.

## Original README

Disclosure: I work on **AgenC Core**, the open-source agent harness used here: [https://github.com/tetsuo-ai/agenc-core](https://github.com/tetsuo-ai/agenc-core)

I gave DeepSeek V4.1 Flash (effort max) one prompt and the original Sopwith executable from 1984: a 43 KB DOS binary, with no source code. It had REA connected as an MCP server ([https://github.com/morluto/rea](https://github.com/morluto/rea)), which runs Ghidra under the hood. Then it worked on its own for an hour, in a sandbox with no internet.

What it did:

* Decompiled the whole binary with REA and worked out the game loop, controls, flight physics, scoring, computer planes and world layout.
* Wrote a Python extractor that pulls everything straight out of the EXE: the sprites (the plane at every angle, buildings, oxen, birds), the 3,000-point terrain profile and the game tables.
* Rebuilt it as a plain HTML/Canvas game at the original 320x200 CGA resolution, with PC-speaker style sound.
* Ran the original in DOSBox, captured reference frames, and compared them side by side with its port in headless Chrome. The only thing not inside the EXE was the BIOS font, so it recovered that from the DOSBox screenshots.

The numbers:

* 1 hour, one prompt, no follow-ups
* 374 model calls, 624 tool calls (7 of them REA)
* 145M input tokens (99.7% cache hits), plus 308K output
* $0.68 total at DeepSeek’s off-peak rate ($1.37 at peak)

The flying in the video is a small script pressing the game’s real keys, so you can see takeoff, the bombing runs and the autopilot landing in one take.

The full prompt is in the comments. The author released Sopwith’s source under the GPL, but the agent never saw it: it worked only from the binary.
