Multi-Agent OS for LLM-powered NPCs — AI characters live, wander, and talk autonomously in a pixel-art 2D world. Supports MCP protocol, skill system, multi-LLM (DeepSeek/OpenAI/Claude), and real-time collaboration chains.

中文 | English

Agent-Worlds

Multi-Agent OS — AI characters live, wander, and talk in a pixel-art world

What is this?

Agent-Worlds is an operating system for AI agents. You drop NPC characters into a 2D pixel-art world, and they:

• Wander around the map autonomously
• Bump into each other and start conversations
• Remember what happened (RAM buffer during chat, persistent HJL memory after)
• Use tools — browse the web, generate images, write code, call external APIs
• Collaborate — chain multiple NPCs to complete complex workflows

Think of it as a SimCity for LLM agents, with a real-time pixel map, a skill system, and MCP protocol support.

Features

• Collision-driven social engine — NPCs meet on the map and autonomously decide to chat
• Pixel-art 2D map — Real-time rendered 48x48 pixel world with multiple scenes
• Skill system — Give an NPC abilities with one line of config: "skills": ["programmer"]
• MCP protocol — Connect external tool servers (browser, database, etc.)
• Layered memory — RAM buffer (live conversation) + HJL file (persistent history)
• Multi-world — Switch between worlds: modern city, wasteland camp, comic studio
• Comic production pipeline — Example: scriptwriter -> storyboard -> artist -> layout -> voice actor
• Multi-LLM support — DeepSeek, Zhipu GLM, Volcano Engine, OpenAI-compatible, local models
• Player mode — Join the world as a player and talk to NPCs directly

Screenshots

NPC Conversation	World Management	NPC Configuration
		![NPC Con