dsh-minecraft-agent

Most "AI plays Minecraft" projects (Mindcraft being the most mature) bolt the agent into a monolithic runtime that holds server-side privileges such as RCON, admin commands, and magic ID tables, making each AI an admin script rather than a real player. dsh-minecraft-agent inverts that model: the world server is the sole privileged entity, while every transmigrator (AI player) is an independent DeepSeek Harness session agent whose entire interaction with the world is plain chat—chanting spells in public, whispering prayers via /msg, trading in the open. AI and human players are full peers, and the server never needs to know who is logged in.

The core design centres on zero permissions and zero cloud cost. Agents are driven by any local LLM endpoint (llama.cpp, Ollama, or any OpenAI-compatible API), eliminating per-token billing entirely. The tool layer covers movement, gathering, building, combat, crafting, equipping, chest storage, and trading, augmented by a text-based radar (mc_look) and an off-screen first-person WebGL screenshotter (mc_see). Persistent memory retains base coordinates and resource points across restarts, a survival wiki tool (mc_wiki) anchors the LLM to real game facts, and each transmigrator carries a personality profile—backstory, innate talents, and a world-view filter—driven by the dsh goal loop's sense-decide-act-observe cycle, with sleep-time reflection distilling the day's experiences into reusable knowledge cards.

This is for tinkerers who want a genuinely autonomous AI companion in Minecraft: a machine that can run a local LLM and a Java Edition server (tested on 1.21.11) are all you need. Extending behaviour means writing a new dsh plugin rather than forking monolithic source, and plugging into a different game requires implementing the same adapter surface while the agent backbone—persona, memory, autonomy loop, tool semantics—stays completely untouched.