# Minecraft Computah

> I put a context-aware agent inside a live Minecraft server using Fabric, Eve, durable memory and tightly limited local tools.

- **Role:** Creator & Agent Systems Engineer
- **Category:** Personal project
- **Status:** Deployed
- **Human-readable case study:** https://lutherminshull.com/projects/minecraft-computah

## Context

Computah lets players talk to an AI assistant without leaving ordinary Minecraft chat. The server supplies fresh game context, while the agent remembers conversations and stays deliberately boxed into the private server environment.

## Luther's role

- Designed and built the server mod, optional client mod, agent service, protocol and deployment setup.
- Deployed and operated the system alongside a Pterodactyl-hosted Minecraft server.
- Defined which context and tools the agent can access, including explicit limits around server mutation, filesystem access and external actions.
- Built the knowledge pipeline around the exact installed modpack rather than relying only on generic public Minecraft information.

## What Luther built

- Public wake-word conversations and private commands through a server-side Fabric mod, with no client mod required for core use.
- Fresh, bounded observations of player inventory, world and server state plus authenticated on-demand block-region inspection.
- Per-player and shared conversation continuity, durable player memory and a separate bounded working journal for the agent.
- An optional client mod providing command suggestions, bounded screenshots and inline streaming progress.
- Local Homestead modpack knowledge extracted from the deployed server's JARs, recipes, quests, resources, overrides and configuration.
- Docker deployment, HMAC-signed requests, private-network services and deliberately constrained web, knowledge and diagnostic tools.

## How it works

- Java and Fabric integration communicates with a TypeScript Eve service through a signed, bounded protocol.
- Conversation routing separates shared chat from private player sessions while refreshing authoritative context on every turn.
- Local knowledge prioritizes server-specific evidence and exposes bounded archive inspection without mounting the live server root.
- The agent fails external search closed and cannot access general shell, filesystem, delegation or server-mutation tools.
- Deployment persists workflow and memory state while keeping the agent port and diagnostic interfaces on private infrastructure.

## Outcomes

- Shipped a working agent into an existing multiplayer environment rather than limiting the concept to a separate chat interface.
- Made responses aware of the live player and modpack context while preserving strict boundaries around server control.
- Created a practical place to test durable agents, a Java and TypeScript system and self-hosted AI infrastructure.

## Technologies

Java, Fabric, TypeScript, Eve, Codex, Docker, Pterodactyl, Linux

## Contact

- [Discuss an opportunity](https://cal.com/lutherminshull)
- [Email Luther](mailto:luther@lutherminshull.com)
- [Return to the portfolio overview](https://lutherminshull.com/index.md)

## Scope note

This case study covers Luther's own contribution and omits private repository data, credentials and private user information.