It is an open-source AI agent I made that runs in a Debian Docker container with Chromium as the default browser. It is model agnostic (LiteLLM) and can do normal agent things like file manipulation, browsing, and coding.
The special thing is its ability to adapt, improve itself, and learn. When it sees that it can't do a task because its toolset won't let it, it creates and saves a new tool. If it hits an error, it reflects on the error and writes how to avoid it in the future in a special JSON file.
You can even make it change its system prompt if you want it to act in a certain way. It also has persistent memory and the ability to schedule tasks.
Of course, whenever it makes a new tool or changes its prompt, it asks for your permission. It also asks whether or not you want it to need approval for it to run the tool it made.
You can talk with it through a beautiful web interface at localhost:5000 and see it work in real-time through noVNC.