I work on several research projects at once, and I needed the agents to do the legwork while I kept track of everything and lead the project.
crux is an open-source toolbox for agentic computational research. It all lives in a local vault in your project repo.
The knowledge base expands Karpathy’s LLM wiki: crux crawls the citations of a few seed papers through OpenAlex, and an LLM compiles them into linked markdown pages the agent can traverse.
The research tree holds the project’s questions and the hypotheses that test them. Each hypothesis has a null and verifiables defined before the run, so its result can’t be p-hacked. The agents grow the tree, and you watch it grow in a local browser view.
Autopilot is an expanded autoresearch loop for questions with a metric to optimize. Each attempt is a hypothesis in the tree, and the next attempt is picked with the search rule from Google’s ERA.
crux is young (the first commit was in June 2026), and contributions are welcome. I’d like to hear how you use agents in your research.
**Website**: [Crux — a scientific-method lab notebook your AI agent drives](https://mehdiforoozandeh.github.io/crux/)
**GitHub**: [https://github.com/mehdiforoozandeh/crux](https://github.com/mehdiforoozandeh/crux)