Claude Science: Anthropic's beta for scientific research

My notes

Summary

Claude Science is a public beta app (not a new model) that turns Claude into a research agent running full scientific analyses on the user’s own infrastructure. It manages compute environments, connects to 60+ scientific databases and domain models, and attaches full provenance (code, environment, conversation) to every artifact so results are reproducible and defensible months later.

Key Insight

• It’s an agentic workbench, not a model. Uses whatever Claude model your plan already includes. The value is everything around it: scientific tools, DB connectors, and compute orchestration.
• Compute runs where your data lives - laptop, lab Linux box, HPC login node, or cloud VM. Writes batch scripts and submits/manages Slurm jobs over SSH, or via a Modal account. Variables, dataframes, and loaded models stay in memory across a whole analysis for fast iteration.
• Provenance is the differentiator. Every figure/table/notebook ships the exact code, environment, plain-language description, and the conversation that produced it. A background reviewer flags incorrect citations, untraceable numbers, and figures that don’t match their code before results surface.
• Domain-ready day one across genomics, single-cell, proteomics, structural biology, cheminformatics. Renders proteins, alignments, genomic tracks, chemical structures, and PDFs natively. Pipelines can be saved as reusable skills that future sessions inherit automatically.
• NVIDIA integration: uses skills from NVIDIA’s BioNeMo Agent Toolkit to connect to BioNeMo models/libraries including Evo 2, Boltz-2, and OpenFold3.
• Data posture: app runs on your infra, raw datasets and compute stay local; only prompt/response content is processed by Anthropic under standard retention.
• Access: beta on macOS and Linux for Pro, Max, Team, Enterprise. A discounted Team plan for research labs exists for academic/nonprofit labs (PI-verified); for-profit and CRO teams go via Team/Enterprise.