Town: Discord meets pixel-art NPCs

Most people are just using LLMs as glorified autocomplete, but this dev actually turned Claude into a living, breathing pixel town. It’s basically Discord, but instead of just typing in channels, you’re walking an avatar through a map and talking to NPCs that actually have agency.

I’ve been playing around with the concept of using sub-agents to stress-test my own logic or act as a sounding board for technical docs, and this project, "Town," takes that workflow and wraps it in a game engine. The engineering behind it is what caught my eye: each NPC isn't just a chatbot; they are scoped agents with specific tool permissions. You can't just ask a barista to run a database query if the dev only gave them "talk" permissions. If it's a detective NPC, they might have a close_case tool. It’s a clever way to manage LLM hallucinations and keep the scope tight.

The whole thing is open source and surprisingly easy to deploy. A "town" is literally just a JSON file and some MDX folders. The identity and system prompts live in the MDX frontmatter. There is even a Claude plugin that supposedly builds the entire environment from a rough description, which sounds like a massive time-saver for rapid prototyping.

Here is what's available if you want to poke at it:

Startup Town: Pitch ideas to AI versions of famous investors. It's basically a sandbox for stress-testing your business plan.

Murder Mystery: A logic game where you have to catch NPCs in contradictions to solve a crime.

Roast Town: For when you just want to get destroyed by an AI.

CORE Town: The demo hub. It's got everything from noir detectives to YC House simulations.

If you want to see the source or build your own, check the links below:

Source: https://github.com/RedPlanetHQ/town

Explore Towns: https://town.getcore.me/explore

It's a wild way to think about community engagement and AI workflows. Instead of a static chat interface, you're interacting with a persistent, agentic world. Definitely worth a look if you're into agentic workflows or just want to see how far you can push LLM tool-use in a real-world (well, pixel-world) scenario.