Remember Lemmings? The tiny green-haired suicide bombers? I spent last weekend rebuilding them with AI. Not a remake — a full reverse-engineered port, complete with Adlib, Tandy, and EGA sound that was never documented. And the result is a glorious, broken mess that only runs on Chrome Canary with a secret flag.
Let me explain why that’s actually the most exciting thing I’ve seen in AI this year.
I’ve been deep in the retro gaming rabbit hole for years. Manual porting is tedious. You stare at disassembly dumps, map memory addresses, pray to the hardware gods. It’s archaeology with a keyboard. But last week, I tried something different. I hooked up Claude Code to a Ghidra MCP (a reverse-engineering tool) and a DOSBox MCP. Then I gave it a simple command: “Reimplement Lemmings. Include the undocumented sound routines.”
It worked. Not perfectly, but scarily well.
“The future of software preservation isn’t manual porting — it’s AI agents acting as automated reverse-engineers, decoding undocumented systems in real time.”
Here’s the kicker: the AI didn’t just copy existing code. It used the disassembler to figure out how the original Adlib and Tandy routines worked, then generated working JavaScript implementations. I gave it zero context on those sound chips. It figured out the register writes from the DOS binary. That’s wild.
But there’s a twist. To make this work, I used the experimental layoutsubtree and drawElementImage canvas APIs. These are bleeding-edge, unstable, and only available in Chrome Canary with a flag enabled. So 99% of you reading this can’t even play the demo. One commenter on Hacker News already slammed it: “Thanks for the code, but I can’t play it.”
And that’s beautiful. Because it highlights the paradox of using cutting-edge AI to meticulously resurrect obsolete hardware, while relying on such fragile browser APIs that the final product breaks for nearly everyone.
You know what? That’s fine. This isn’t a consumer product. It’s a blueprint.
“Neutrality is death. This is brilliant — not because it works, but because it proves a new method.”
I’ve seen the comments: “Why not just use DOSBox directly?” Because that’s not the point. The point is that AI can now autonomously navigate niche reverse-engineering tasks. It’s not about making Lemmings playable in a browser. It’s about proving that an AI agent can hook into a disassembler, understand undocumented hardware, and generate working code — all without human intervention.
This is the future of software preservation. You want to save a 1995 game that uses a custom sound chip? You don’t need a team of reverse-engineers. You need an AI agent with the right MCPs.
And yes, it currently only works for 1% of users. But so did the first web browsers. So did the first AI models. The technology is raw, but the direction is clear.
So here’s my challenge: stop porting games manually. Stop treating software preservation as a museum curation exercise. It’s a machine learning problem. Let the AI do the dirty work. We’ll handle the creativity.
“The best articles make you rethink something you thought you knew. I thought reverse-engineering required human intuition. Turns out, AI can be just as obsessive — and twice as fast.”
What’s the cost? A few hours of setup, and a willingness to embrace broken experimental APIs. But the payoff is a new paradigm: AI-assisted archaeology. I’ve already started on a second project — a 1997 DOS game with a proprietary audio driver. Claude Code is now generating the port. And I’m not touching a single hex editor.
If you’re a developer, an archivist, or just a nostalgia junkie, this is your wake-up call. The tools are here. The workflow works. The only thing missing is your willingness to let go of the manual approach.
Or, you know, you can wait for Chrome to ship the stable API. Either way, the future is already running — just not on your browser yet.
FAQ
Q: Why does the demo only work in Chrome Canary?
A: Because it uses the experimental `drawElementImage` and `layoutsubtree` canvas APIs, which are not yet shipped in stable Chrome. This is a deliberate choice to showcase the cutting-edge capability of AI-driven reverse-engineering, not a production-ready product. Expect stable support within 6–12 months.
Q: Can I use this workflow for my own retro game projects?
A: Yes, with caveats. You need Claude Code (or another LLM with MCP support), a Ghidra MCP server, and a DOSBox MCP server. The setup is non-trivial and requires some familiarity with reverse-engineering tools. But the core principle is proven: you can feed an AI a disassembler and get working code. Start with simple games and expect buggy output.
Q: Isn't this just a fancy way to play Lemmings in a browser? Why not use DOSBox?
A: That's missing the point. This isn't about playing Lemmings — it's about proving that AI can autonomously reverse-engineer undocumented hardware. DOSBox emulates the original environment; this workflow extracts the logic and rebuilds it natively. That matters for software preservation, porting to modern platforms, and even for creating educational resources. It's a method, not a product.