Bot Aquarium is an experimental open-source agent that hands control of a Linux machine to a large language model and lets it pursue a stated goal by issuing real terminal commands. Its tagline is “agents before they were cool”, and the project predates much of today’s agent tooling. Written in Go and released under the GPL-3.0 license, it has about 660 stars on GitHub and was last updated in August 2025.
The loop is simple and revealing: you pass a goal such as “run a Minecraft server”, the model answers with one Linux command, the host executes it, and the terminal output is summarized and fed back into the next prompt. In the documented example the agent updates packages and installs a Java runtime step by step toward its goal. Aquarium can talk to OpenAI’s models or to a local model served through llama-cpp-python, which keeps experiments cheap and private. Two context modes control how much terminal history the model sees: a partial mode with the last lines of output, or a full mode with the entire transcript. A Docker-based setup keeps the experiment contained on its own network.
Bot Aquarium is free and open source under the GPL-3.0 license. If you use a hosted model you pay that provider’s API rates; local models run at no extra cost.
Bot Aquarium is for AI tinkerers and agent researchers who want a small, readable codebase that demonstrates the core agent loop of planning, acting and observing on a real machine.