ScienceClaw

ScienceClaw is an open source AI for Science agent that improves from verified execution. Each task is solved as a typed, executable workflow, and when a repair is reproduced under a clean replay it becomes a linked Skill and Operator, kept only if independent validation tasks improve. It is aimed at researchers and teams building scientific agents. What evolves is a versioned program of Skills, covering decomposition, workflow construction and recovery, and typed Operators with explicit input, output and domain contracts. Ports carry schemas for type, shape, unit and provenance, and nodes are fingerprinted so an edit reruns only what it affects. Results are regenerated from a reset environment and checked against hard scientific constraints. An update persists only if replay reproduces and uses it, every hard check passes on independent validation tasks, cost stays within budget and the score strictly improves. Updates are versioned snapshots with receipts, and in the gateway they wait as candidates until a user promotes them, with rollback to any version. The project reports evaluation across 23 disciplines, a 16.45 percent mean out of distribution gain over the frozen agent and a 98.23 percent hard constraint pass rate. The site links to the code on GitHub, a quick start, a paper on arXiv and benchmark data. No pricing is published.