Voyager is an open-source, LLM-powered embodied agent that plays Minecraft autonomously — exploring the world, learning new skills and making discoveries indefinitely without human intervention. Built by the MineDojo team with NVIDIA researchers, it has around 7,300 stars on GitHub.
Voyager was the first lifelong-learning agent of its kind. Instead of being trained on fixed tasks, it is driven by GPT-4 through three components: an automatic curriculum that proposes ever-harder exploration goals based on what the agent has seen, an ever-growing skill library of executable code that Voyager writes for itself and reuses for more complex behaviors, and an iterative prompting loop that folds environment feedback, execution errors and self-verification into the next attempt. The result is an agent that learns to chop wood, build tools, fight mobs and unlock the Minecraft tech tree far faster than earlier approaches, and can even apply its learned skills in fresh worlds. The code and paper artifacts are published for research under the MIT license; the project was last updated in 2024.
Voyager is free and open source under the MIT license. Running it requires your own OpenAI API key for GPT-4 calls and a Minecraft environment, but the agent code itself costs nothing.
Voyager is for agent researchers, game-AI enthusiasts and developers studying autonomous, self-improving systems — and anyone who wants a concrete, well-documented example of an LLM agent that learns from experience.