AgentJam
← Home
Minecraft Embodied Agent Voice Education

AgentJam

AI Mentors That Play Minecraft With You

The mentor is the headline. It plays Minecraft with you, not instead of you.

Players Supported

1–6

Model

DeepSeek

Voice

Spatial (fades w/ distance)

Platform

Minecraft 1.21.x

AgentJam puts AI mentors inside Minecraft as real players — not chatbots, not overlays, not passive tutors. The agent walks, mines, builds, and fights in the actual Minecraft Java client. It speaks lines out loud with spatial voice that fades with distance. It plays alongside students as a peer and mentor, turning the game itself into the classroom.

In teaching mode, AgentJam mentors individual players through tasks. In autonomous classroom mode, it runs 50–60 minute spoken-English sessions for 4–6 students — directing the flow, scoring performance, and orchestrating peer AI bots that model the target language. Every interaction happens inside the world, not beside it.

Features

Real Player-Bot

In-World

Walks, mines, builds, and fights in actual Minecraft Java client. Joins as a real player via the Minecraft protocol — no mods required on the student's side.

Voice Communication

Spatial Audio

Speaks lines out loud with volume that fades naturally with distance. The mentor's voice is anchored at its position in the world — walk away and it gets quieter.

Teaching Mode

Plays With You

Plays alongside you as a mentor, not a passive tutor. Demonstrates tasks, guides exploration, and responds to questions in real time — inside the game.

Autonomous Sessions

50–60 min

Fully autonomous spoken-English classes for 4–6 students. The AI captain directs the flow, peer bots model the target language, and every student gets live scoring.

Architecture

Mineflayer → LLM → Spatial Voice. The agent runs as an independent process, joins as a real player via the Minecraft protocol, and produces spatial audio anchored at its in-world position.

Game Layer

Mineflayer

Connects to Minecraft Java 1.21.x as a real player. Full control: movement, block placement, combat, inventory. Observes world state, player positions, and chat.

Mineflayer 1.21.x Real Player

Reasoning

DeepSeek

LLM processes world observations, player actions, and conversation context. Decides what to do next — move, build, mine, speak. Runs with tool-use for Minecraft actions.

DeepSeek Tool Use Context-Aware

Voice Output

Kitten TTS + Spatial

Streaming TTS with spatial audio. Voice volume and stereo panning computed from agent-to-player distance. Multiple voices per session for peer bots and mentor.

Kitten TTS Spatial Audio Multi-Voice
Minecraft Protocol DeepSeek Reasoning Spatial TTS Open Source

Demos

Papers & Reports

Technical reports and papers. All work is open access with accompanying code.

AgentJam: AI Mentors That Play Minecraft With You

Technical Report · 2025

Read ↗

Spoken English Sessions: Autonomous Multi-Student AI Classes in Minecraft

Technical Report · 2025

Read ↗