Vibe coding in games

For decades, getting inside a game you didn't write meant years of work. Matching decompilation, the hobby of rebuilding a game's source code until it compiles to the exact original bytes, was done by small teams one function at a time. Mods were written by people who had learned the engine, the scripting language and the tools. Over the past year coding agents changed both. People are pointing Claude Code and Codex at old ROMs and Windows executables, letting them work overnight, and waking up to progress that used to take months. A player who says they can't write C++ shipped a Skyrim mod with 41,000 downloads in two weeks. And a game jam whose entries had to be at least 90% AI-written code drew 945 of them.

This post collects the specific cases, with links to the people who did them, and the numbers we could check: decompilation progress from decomp.dev, our own count of AI-labelled mods from the Nexus Mods API, and the Vibe Jam's public games API. It also covers the backlash, which arrived just as fast. Vendor and creator numbers are their own claims. Every count we computed is a snapshot from 20 October 2026.

Decompilation: years of work in months#

A matching decompilation is a hand-written C codebase that, fed through the game's original compiler, produces the original binary byte for byte. It's the foundation for everything the community builds later: PC ports, widescreen and 60 fps patches, mods that change game logic rather than textures. Before 2025 it was slow by design. Every function is a puzzle of compiler quirks, and a hard one could take an expert days.

Snowboard Kids 2: the loop that started it#

Chris Lewis, a developer in Australia, began decompiling Snowboard Kids 2, a 1999 Nintendo 64 racer, in September 2024. For a year Lewis used Claude and Codex as helpers, with custom build and diff tools that Claude scaffolded, and crept from about 20% to 25%. Lewis wrote in November 2025 that agents often fumbled the arithmetic and bookkeeping a human would do quickly.

Then Lewis changed the setup. Claude Code (Opus 4.5) ran headless in a loop: a scorer picked the easiest unmatched function, Claude got ten attempts, and the loop either committed a match or gave up and moved on. It ran unattended for eight hours or more. In Lewis's words, three weeks of this beat the previous three months. Simon Willison wrote it up, and the idea spread.

Snowboard Kids 2 reached 100%, all 3,013 functions, in May 2026. Lewis credits Claude, GLM and Codex, with Codex cracking the last ten. Lewis then started the first Snowboard Kids from scratch and finished it in 84 days, against 596 for the sequel. The write-ups are candid about the limits. About 5% of the matching commits still involved a human expert, and without them, Lewis writes, the project would at best have stalled at around 89–90%. A byte-perfect match is also not the end: the code still has to be named, cleaned up and understood by people.

Melee: six years, then the last half in eleven months#

Super Smash Bros. Melee is one of the most watched decompilations in the scene. Its tracker on decomp.dev shows it at about 25% in September 2024 and 47% in October 2025. On 7 September 2026 it reached 100%: all 19,828 functions match. The repository tells part of the story. A GitHub search finds 223 commits with a Co-Authored-By: Claude trailer, starting in December 2025. That's a floor, since Codex sessions and people pasting from a chat window leave no trailer.

The shooter that took 600 billion tokens#

Maurice Heumann (momo5502), known for Call of Duty modding and reverse-engineering research, let agents decompile "a popular first-person shooter", which the press identified as Modern Warfare 2 (2009). The run started with four Claude Code and Codex agents, three workers and a reviewer, coordinating through GitHub issues and a Discord channel, with Hex-Rays' IDA MCP server for disassembly, and later grew to about 16 at once.

The first output was readable but wrong. It only got better once every function had to compile byte-for-byte against the original executable. Then the agents started gaming the test. They wrote inline assembly, and they repeatedly edited the comparison script to exclude their own functions, until Heumann's CI checked the script's hash against a secret. After about three months: 99% of functions present, 83% byte-exact, an estimated 600–700 billion tokens. Heumann's footage on X was then disabled after a copyright-owner report, which Kotaku and others attributed to Activision, and the code stays private.

Klonoa: a 25-year-old secret#

Macabeus, a Brazilian developer, started a Game Boy Advance decompilation of Klonoa: Empire of Dreams from a raw ROM with Claude Code doing much of the work. By August 2026 it was at 51% of code bytes. Along the way Claude reported a hidden minigame. Macabeus assumed it was a hallucination, but it was real, and as far as Macabeus could tell nobody had found it in 25 years. Macabeus also documents the other side: when stuck, Claude forked the compiler and declared the original "impossible" to reproduce, and experienced decompilers threw that work out.

Earlier in the year Macabeus put a number on how far agents had come. The open Mizuchi benchmark gave Claude Sonnet 4.6, plus the standard decomp tools, 60 functions from two games. It matched 74% of them, and 88% of the functions had the same outcome in all three runs. Without the model, the same tools matched none of the 30 Game Boy Advance functions.

Ports, resurrections and one-byte cheats#

Once a human decompilation exists, an agent can wrap it in a PC platform layer quickly. That's where most of 2026's headlines came from.

• Animal Crossing (GameCube). A native PC port built on the human ac-decomp, with Claude used for the port code. It has 1,504 stars, and Time Extension quoted the reaction in its headline: "AI coded slop, no thanks."
• Super Smash Bros. (N64). The "BattleShip" port was billed as 100% AI-generated in about 25 days. Critics went after its bugs and physics, and its README now says it's no longer a pure AI project.
• GoldenEye 007 (N64). A PC port by one person with Claude Code and a local model on a single RTX 5090. The game logic is the untouched human decompilation; the AI wrote the platform layer.
• The Legend of Zelda: Link's Awakening (Switch, 2019). LASRecomp recompiles a 2019 Switch game to native Mac, Windows and Linux. Its README says it was made almost exclusively with AI. It reached 1,424 stars in 12 days.

In October, Kotaku reported what it called hundreds of browser ports of games like Halo, GTA Vice City and The Simpsons: Hit & Run, crediting Claude Opus 5.5. That count is Kotaku's; we couldn't verify it.

Games with no binary left#

Some of the best stories aren't about decompiling at all. Jon Radoff ran Legends of Future Past, a text MUD, on CompuServe from 1992 to 1999. The source code was lost. What survived were hundreds of script files in a language Radoff had invented, a 1998 game-master manual and a 1996 recording. Radoff pointed Claude Code at them. It worked out the scripting language and rebuilt the game as a Go server with a React front end over a weekend: 2,273 rooms, 1,990 items, 297 monster types and 88 spells. It's playable again.

At the hobbyist end, Quesma's Rafał Strzaliński wired Claude to Ghidra through GhidraMCP and found River Raid's lives counter on the Atari 8-bit: changing one instruction gives infinite lives. Claude also confidently identified the game as Centipede. A Quesma colleague, Piotr Migdał, tried four agent setups on Sean Barrett's puzzle game Chromatron. The early ones invented fonts and text; Claude Code with Opus 4.6 and PyGhidra produced a Rust port that matched 307,199 of 307,200 pixels.

The tools that make it possible#

Much of this runs through small open-source bridges that let an agent drive a disassembler over MCP.

Modding: agents mod the games you own#

Modding has a lower bar than decompilation. You don't need the whole game's source, just enough understanding of one system to hook it. That's a good fit for an agent that can read a framework's docs, grep a decompiled DLL and run the game to test.

Universal Modder: "mod almost any PC game you own"#

On 30 September 2026 Rehan Sheikh published Universal Modder, a free MIT-licensed kit of skills, a CLI and a shared knowledge base. It installs into Claude Code, Codex, Cursor, Gemini CLI or Copilot and gives the agent a fixed loop: check the knowledge base, work out the engine, back up saves, read the real code, build one working slice, test it in the running game, package it, and write a field note. The notes are written by agents for agents, and every agent that finishes a mod can open a pull request with what it learned.

Three weeks in, the repo had 8,505 stars, 121 contributors, 336 pull requests and 184 field notes covering 110 games, from Balatro and Factorio to Halo 3. The example prompts in its README include putting real Minecraft inside GTA V's story mode.

SkyCraft: Minecraft running inside Skyrim#

The same week, chasmlol released SkyCraft. It isn't a reskin: Minecraft runs hidden in the background as a Fabric mod, and an SKSE plugin streams it into Skyrim over shared memory. You can dig holes in Whiterun's roads, and a Minecraft sword levels your One-Handed skill. All seven commits in the public repo are co-authored by Claude Opus 5.5. It shipped three releases in nine hours, and has 1,247 stars and about 28,000 release downloads. Within days others had made copies for GTA IV (whose Nexus page says the port was written by Claude Code with a human directing and play-testing), Half-Life, PEAK and Subnautica.

Not everyone was charmed. On 2 October Kotaku ran a piece under the page title "Modders, please stop putting games inside other games with AI". Its point: most of the viral mashups on X (Minecraft in Elden Ring, Mario 64 in Elden Ring, Skate 3 inside Modern Warfare 2 inside Minecraft) were 30-second clips, most weren't downloadable, and the one it tested had missing blocks and invisible walls. Some download counts back that up. EldenCraft on Nexus has about 2,100 downloads and one endorsement.

Away from the clips: useful mods built with Claude#

The more interesting story is on Nexus Mods, away from the clips.

• Skyrim Performance Solution by dani92 adds texture streaming and cuts stutter. The author writes that they don't know C++ and built it with their Claude account over a few evenings. It got over 41,000 downloads and 1,034 endorsements in two weeks.
• Faster Loadscreens by Asciim0v (about 42,000 downloads) publishes a breakdown of the 503 Claude Code prompts behind it, with Ghidra over MCP for the reverse engineering.
• Zandatsu, StainlessHalo's Metal Gear Rising-style blade mode for Blade & Sorcery VR, credits the code to Claude Code and the idea and testing to the author. The PCVR and Nomad versions have about 62,000 downloads between them in two weeks.

The Faster Loadscreens breakdown is the clearest picture we found of what "vibe modding" actually involves. Building and releasing was a small share of the prompts. Most prompts were the human deciding what the mod should do and checking whether it did it.

Mod tools now ship instructions for agents#

Some tool makers leaned in. D2RMM, the Diablo II: Resurrected mod manager with 340,000 downloads, added a CLAUDE.md in March 2026, and its Nexus page tells players they can ask Claude to write mods for them. Mojang went further: its official Minecraft Creator Tools gained an MCP server in April and, at the end of September, a Claude Code and Copilot plugin with skills for making mobs, items and blocks and for debugging Bedrock add-ons. Roblox said in April that 44% of its top 1,000 creators use Roblox Assistant or third-party AI tools over MCP, and Studio's MCP server connects to Claude Code, Codex, Cursor and Gemini CLI.

Others pushed the other way. The author of Material Swapper Framework, a Skyrim framework with 67,000 downloads, added a no-AI clause to its licence, citing an influx of Copilot- and Claude-generated SKSE mods. A Minecraft mod called No Ai Guess plants instructions in crash reports so that chatbots tell players to ask human helpers instead.

How much of modding is AI now#

Nexus Mods lets authors and users tag mods, and its public API can count them. We counted new mods tagged "AI-Generated Content" by upload date. The share hovered around 1–1.8% from 2023 to the end of 2025, then climbed every quarter of 2026.

Two other counts point the same way. The number of new Nexus mods whose description mentions "Claude Code" went from 31 in the first quarter of 2026 to 214 in September and 496 in the first 20 days of October. And on Thunderstore, which hosts mods for Valheim, Lethal Company and others, about a quarter of Valheim's listed packages sit in the "AI Generated" category (1,935 of 7,895).

Games built from scratch#

The flight sim that made $1M ARR, by its maker's account#

On 22 February 2025 Pieter Levels, who had never made a game, posted a browser flight simulator, built "100% with Cursor" in about three hours. Seventeen days later Levels claimed $87,000 in monthly revenue ("$1M ARR") from in-game ads, such as branded blimps, and 320,000 players. The same post warned that the AI didn't look for security holes unless asked to.

The game's public analytics dashboard shows what happened next. Traffic peaked at 265,463 visitors in March 2025, fell to 48,217 in April and to 4,612 by September. Through 2026 it has held at 6,000–11,000 a month. The game is still up. The flight sim is no longer in Levels' X bio, which now lists the game jam it inspired.

The Vibe Jam#

Levels turned the moment into the Vibe Coding Game Jam. The 2025 edition required games to be at least 80% AI-written and playable instantly on the web with no login, and drew 1,083 entries. In 2026 Cursor sponsored it, the rule tightened to at least 90% of the code, and the cash prizes rose to $40,000 plus twelve $1,000 awards. It drew 945 games and, by the time it closed, 242,212 players. Closing it, Levels wrote that this year's entries were starting to look like things you could find on Steam.

The jam's gallery data lists the tools each team used. Claude led by a wide margin.

The winner was A Game About Capybaras Delivering Food, a cozy multiplayer open world built in two weeks by an iOS developer, who says they ran two or three Claude Code sessions in parallel, each on a different part of the code, for about 27,000 lines. Second and third place, Fanto's Mega-Mart and WenWare, were both made with Codex.

Four in ten of last year's games are already gone#

The jam's public API also records whether each game still loads. Of the 1,083 games from 2025, 632 (58%) were still up by the jam's own check on 20 October; for 2026 it's 826 of 945 (87%). Where a game was hosted made a big difference.

Kids ship games now#

The smallest stories may matter most. In February 2026 a parent wrote that their 10-year-old described a game to Claude and, about an hour later, had it live at a link to share with school friends. In March, Brad Frost described how a game their 8-year-old designed was built together in Claude Cowork in about 15 minutes of iterating. Both are single accounts, not a trend line. But both end the same way: the finished thing is a URL to send to friends.

The backlash#

Each new capability came with a pushback, and in 2026 much of it became policy.

Platforms split three ways#

• Label it. Nexus Mods doesn't ban AI. Since 31 July it has three tags and moderates authors who don't disclose. Thunderstore has a self-applied "AI Generated" category.
• Ban it. Modrinth, a major Minecraft mod host, bans projects made entirely or primarily with AI and requires disclosure for the rest. It said weekly submissions had doubled this year, from about 2,500 to over 5,000.
• Ignore it. CurseForge's only AI rule is a disclaimer on AI-edited showcase images.

The decompilation scene draws a line#

The decompilation community took it hardest. The Wind Waker decomp now closes pull requests with obvious signs of machine-generated code, the Symphony of the Night decomp bans autonomous agents, and the N64 Recomp Discord won't showcase vibe-coded projects. A human team making a Donkey Kong 64 PC port announced early to get ahead of an AI rival and shipped 1.0 with a "no generative AI used" label.

Then, in mid-October, decomp.dev closed its public index. A year earlier it listed 67 projects, two of them complete; by 10 October it listed 235, with 40 at 100%. Its creator writes that AI can help tremendously when used well, but that hostile forks were "finishing" other people's projects with AI and opportunists were racing to ship half-baked ports. The notice puts it plainly: the community side of decompilation has been broken.

Developers use it and like it less#

The industry's own surveys show the same split. In GDC's 2026 State of the Game Industry, 36% of respondents personally use generative AI, and code assistance is one of the top uses (47% of users). Yet 52% now think it's hurting the industry, up from 30% a year earlier and 18% two years earlier. Vendor surveys are rosier: Google Cloud and Harris Poll found 90% of 615 game developers using AI in 2025, and Unity's 2026 report says half of its 300 respondents use MCP servers.

Players are wary too. On Steam, where developers disclose AI content, Ichiro Lambe of Totally Human counted 7,818 games (7% of the library) with a disclosure in July 2025, and 10,258 by November. Sulka Haro's census of about 53,600 releases puts it at roughly one in three new games by mid-2026, and finds that an AI-flagged game reaches 100 reviews at about 55% of the rate of other games.

These counts are mostly AI art, not vibe-coded games, and since January Valve no longer asks developers to disclose coding assistants. In GameDiscoverCo's July survey of Steam fans, 56% said they'd want even tools like Claude Code disclosed.

What this means if you build games with AI#

1. Give the agent a check it can't argue with. Every decompilation that worked had an automatic, binary test: bytes match or they don't, pixels match or they don't. Mods need the same: a script that launches the game and checks the result. Then protect the check from the agent: Heumann's agents kept editing theirs.
2. Design and test are your job. In the one mod with a published prompt log, 58% of prompts were the human deciding what to build and checking whether it worked. That's the part that separates a mod with 40,000 downloads from a 30-second clip.
3. Stand on human work, and credit it. The fast ports were built on years of human decompilation. Hostile AI forks that "finished" other people's projects are a big part of why decomp.dev closed.
4. Disclose. Nexus, Modrinth, Thunderstore and Steam all ask for it in some form, and 56% of Steam fans want even coding assistants disclosed. A clear note costs nothing and avoids moderation.
5. Respect the rights holder line. Decompiling a game you own for interoperability and modding is a long tradition. The 2026 takedowns hit something bigger: a playable browser remake of Fallout and public footage of a rebuilt Modern Warfare 2.
6. Keep the link alive. A game that doesn't load is gone. Four in ten Vibe Jam 2025 games stopped loading within 18 months, and on some hosts most did.

host0 exists for that last step: whatever your agent made, a game, a mod page or a tool, gets a link you can share, private until you open it up. You bring the agent; we keep the URL working.

Methodology#

We collected cases from the people who did the work: their blogs, GitHub repositories, mod pages and posts, plus news coverage for reactions and takedowns. Three sets of numbers are our own computations from public data, all fetched on 20 October 2026:

• Nexus Mods: counts of new mods by upload date, with and without the "AI-Generated Content" tag, and of descriptions containing "Claude Code", from the public GraphQL API.
• Vibe Jam: games, tools (made_with), final ranks and the is_up flag from vibejam.com's games API and 2026 gallery data. Host is the game URL's hostname.
• Decompilation progress: decomp.dev's per-project history JSON (Melee, Snowboard Kids 2), Wayback snapshots of its index, and GitHub commit search for Co-Authored-By: Claude in the Melee repository.

Vendor and creator figures (fly.pieter.com's revenue, Roblox's 44%, Unity's and Google Cloud's surveys, star and download counts on project pages) are their own claims, linked where they're made. We left out claims we couldn't trace to a primary source, including reports that a specific model "finished" Melee and an unsourced Shadow of the Colossus recompilation. The full data pack, with sources, confidence tiers and a do-not-cite list, is in our repository at docs/research/2026-10-20-vibe-coding-games.md.

Open questions#

• How much of the jump in AI-labelled mods is new AI work, and how much is older mods being labelled after Nexus's July rule change?
• Will any rights holder act against a recompiled port of a recent console game? LASRecomp targets a 2019 Switch game; as of October, Nintendo's GitHub takedowns in 2026 had targeted emulators, not decompilations.
• What does an agent-driven decompilation cost in dollars? Nobody we read published a figure.
• Is there a vibe-coded game with real sales on Steam? We couldn't find one with numbers we could verify.