Claude Science pulls a researcher's tools into one workbench

Anthropic has launched Claude Science, a workbench that tries to pull a scientist's scattered tools into one place. The problem is familiar to anyone who does computational research: you bounce between PubMed, Jupyter notebooks, R, databases, and cluster terminals all day. Claude Science puts an agent in the middle of that, with more than 60 pre-configured skills and connectors to resources like UniProt, PDB, Ensembl, and ChEMBL, plus NVIDIA's BioNeMo toolkit. It can render 3D protein structures, genome browser tracks, and chemical diagrams, and produce publication-ready figures next to the code that made them.

The part that matters most for science is reproducibility. Anthropic says every output carries an auditable record of how it was made: the exact code, the runtime environment, a plain-language description, and the full message history, so a result can be validated and rebuilt months later. A reviewer agent also flags citation errors and inconsistencies. The system manages compute from a laptop up to hundreds of GPUs and can process large datasets locally without sending sensitive data out. Anthropic points to early users, including a team at the Allen Institute that built a multi-agent system to write 100-page literature reviews in months rather than two years. It is in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans, and the announcement is here.

Why it matters

If you run wet-lab or computational research, the auditable-artifact design is the feature to scrutinize, because an assistant that records exactly how each figure and number was produced is the difference between a time-saver you can publish from and one you cannot trust.