an agentic lab notebook

AI does the legwork.
You stay the lead scientist.

Projects now grow faster than anyone can keep track. Crux records and organizes the research session — a science tree, a literature wiki, a task list and an autopilot search, in plain markdown, filled by agents while you talk science.

Read the blog post →

scroll

Install

Tell your agent to set crux up in your repo.

npx skills add mehdiforoozandeh/crux --all

Already have months of results? Point it at the repo — it reads what was asked, tested and found, drafts the tree, and waits for your approval before writing anything.

Or open the bundled example first. No agent, no dependencies.

git clone https://github.com/mehdiforoozandeh/crux cd crux && ./crux serve --dir skills/crux/examples/sortlab_vault

Under the hood

Science tree: Questions and falsifiable hypotheses in plain markdown. You set the check; the notebook records the finding.
Literature wiki: Prior work, compiled from sources you curate. It informs the questions. Your findings never flow back.
Taskhub: The work of the lab, as a hierarchy — plans break into subtasks, and a finished experiment waits for your acceptance. Agents drive it.
Autopilot: A search loop for questions that come down to one metric — agents edit the code, run it and score it, under a plan you approve. Every attempt lands on the tree with its verdict.
Skills: Installs into your agent — Claude Code, Cursor, Codex, Windsurf.
Engine: Stdlib-only Python. Nothing is written outside cruxvault/.
Obsidian: The vault is plain markdown with wikilinks — open the same folder in Obsidian and the tree is a graph.