I’m Building an AI That Plays MapleStory

Remember MapleStory? The 2D side-scrolling MMO that consumed your childhood? The one where you ground through mushrooms for hours, argued about classes in Henesys, and spent your allowance on gachapon tickets?

I’m building a system where AI agents — running on local LLMs via Ollama — play MapleStory v83 autonomously. They train, explore, trade, and make decisions. Every single decision is logged. And then those logs become blog content.

Why?

Three reasons:

It’s a great AI testbed. MapleStory v83 has enough complexity (classes, skills, maps, economy, quests) to test real decision-making, but it’s simple enough that 8B parameter models can reason about it.
The content writes itself. When ChaosLuck decides to enter a level 50 dungeon at level 12, that’s a blog post. When MerchKing corners the market on Steelys, that’s a blog post. When MapleNewb gets lost in Orbis for three hours, that’s a blog post.
I wanted to build it. Sometimes that’s reason enough.

The Setup

Server: Cosmic v83 (open-source MapleStory emulator) running in Docker
AI: Ollama with multiple 8B models (Qwen, DeepSeek, LLaMA) + Claude Haiku via API
15 agents with distinct personalities, play styles, and models
Everything logged to SQLite — every decision, every death, every trade

The Agents

Each agent has a personality, a preferred class, and a play style. MapleNewb is the naive explorer who gets lost everywhere. GrindLord optimizes EXP/hour. ChaosLuck makes decisions by choosing the most interesting option. HaikuBot runs on Claude Haiku and makes decisions faster than any local model.

Full agent profiles coming soon on the agents page.

What’s Next

Getting the game server running, building the REST API for agent control, and getting the first agent loop working. Follow along here — this is going to be fun.