I've spent the last several months building an AI-driven system that generates Minecraft mods and plugins e2e. It's an agentic workflow that writes code, runs it through the standard toolchain, and handles the usual compile-time and simple runtime issues as part of the loop. The interesting part has been making the whole process reliable across multiple modding ecosystems and version targets.

The system currently supports generating server plugins (Spigot), server and client mods (Fabric, NeoForge), and Cobblemon mods for Fabric. At this point the system has produced several hundred mods, many of which are running in real production environments. What's been interesting is how people actually use it: about 52% of users who generate a mod come back to create more, and 42% iterate on their mods through conversational refinement with the AI. Roughly 97% of generated mods compile and run successfully on the first pass. That said, 'successful' doesn't always mean pixel-perfect alignment with the user's initial request. The model sometimes drifts near the edges of the spec, especially when the request sits outside a clear operational boundary.

The broader goal is to democratize access to building Minecraft experiences, client-side, server-side, or hybrid. Today, creating custom mechanics often requires either deep familiarity with modding toolchains or paying hundreds or thousands of dollars for bespoke plugins and long-term developer contracts. I want people to be able to rapidly prototype ideas, experiment, and ship things that would've been too expensive, too slow, or simply inaccessible before.

I'm interested in feedback on the approach, the architecture, and the failure modes people expect as requests grow more complex or cross modloader boundaries.