OSS v63 MS Private Server
A MapleStory v63 private server built on open-source foundations — a vanilla, pre-Big-Bang world expanded with a large amount of new maps and content. The game I wished they had kept making. Later, the sandbox where I had seven LLMs learn to play it.
What it is
A MapleStory v63 private server, built on open-source foundations and run in my homelab — not a stock server with the defaults left on, but rebuilt into the game I actually wanted: vanilla, pre-Big-Bang, then expanded. New maps and content, reworked equips and progression, a retuned cash shop, and a great deal of balance tuning, all layered onto that classic world.
The reason
I quit MapleStory after Big Bang. The world I grew up in got flattened into something faster and, to me, hollow — it felt disingenuous. This server was me building the direction I wished the game had taken instead: keep the slow, deliberate, hand-made world, and grow it rather than reset it. A real part of the work was content creation — adding maps and areas the original never had, in the spirit of the old game.
I will be honest about the rest of it: most of the build was a long grind — bandaid-fixing an old client, tuning numbers until they felt right, and chasing bugs that bred more bugs. The reward was getting to stand in a world shaped the way I thought it should be.
Then I let the models play it
Once the server was stable it became a sandbox for a different experiment: seven LLMs playing the live game at once — Opus 4.6, GPT-5.3, Gemini 3.1, Grok 4, Kimi K2, Qwen3, and Sonnet 3.7, each driving its own character. They read game state through packet and memory introspection and acted through the client, which meant reverse-engineering the movement and combat paths to expose them cleanly. I pre-loaded the full context — maps, mob tables, progression — so the models could plan instead of grope blindly. The interesting work was the context engineering, not the agent loop.
The honest ending: seven frontier models grinding a 2D MMO is a beautiful, expensive way to learn that a dumb bot would have been far more effective. So I retired them into plain server-side grinding bots, looping a hunting ground forever with no inference required. Reach for the model where judgment matters, and a bot where it does not.
Status
Done and retired. The world stands; the bots still grind.