MC AI Bot
A Minecraft bot that plans, remembers and works with other bots.
Why it exists
01Minecraft is a good testbed for agents because the world has clear rules, a visible state and failure that is easy to see. If a bot gets stuck, you know exactly where.
I wanted to test memory and cooperation: can two bots share a plan without a human typing every step?
Recorded 2026
Kaushambi, India
The constraints
02- 01Java Edition client behaviour, with Bedrock support through a proxy.
- 02Server plugins and a local runtime rather than a hosted service.
- 03No budget for a large model running continuously.
The approach
03Memory outside the prompt
The bot writes locations, resources and failed attempts to a small memory store. The prompt gets a summary, not the whole log.
Skills as plugins
Mining, crafting and movement are separate modules with explicit preconditions. The model chooses a skill; the skill checks whether it can run.
What it actually does
04- 01Autonomous movement and resource gathering.
- 02Persistent memory of locations and failures.
- 03Plugin-based skills.
- 04Java and Bedrock connectivity through ViaProxy.
- 05Support for multiple bot instances.
Technical notes
05The runtime is Java, with ViaProxy and ViaBedrock bridging protocol versions. A memory store keeps facts between sessions, and an LLM chooses high-level goals rather than individual key presses.
Pathfinding is the bottleneck. The bot can decide to mine iron and still walk into a ravine.
- Role
- Bot runtime, memory design, tool integration
- Year
- 2026
- Platform
- Windows, Linux
- Stack
- Java, Minecraft, ViaProxy, ViaBedrock, LLM APIs
- Status
- In progress
Honest limits
06- 01Still a testbed; it gets stuck in caves and complex terrain.
- 02Inventory management is fragile.
- 03The language model can choose impossible goals.
- 04Not a public server plugin.
What it taught me
07“Embodied agents fail on the boring parts: pathfinding, inventory and error recovery. The interesting reasoning is a small fraction of the work.”