Claude Science

Overview

Freshness note: AI products change rapidly. This profile is a point-in-time snapshot last verified on July 10, 2026.

Claude Science is Anthropic’s dedicated scientific workbench, released in public beta for researchers using Claude Pro, Max, Team, or Enterprise. It is an app rather than a new model: the value comes from bringing literature, scientific databases, code, compute, and publication-oriented artifacts into one environment.

The initial product is especially strong in life-science workflows, but the underlying shape is broader. Researchers can run it on macOS 13 or later or on x64 Linux, connect it to a remote machine or HPC login node, and use Claude as the coordinating layer around existing scientific tools rather than moving the entire workflow into a generic chat window.

Key Features

Claude Science combines a general coordinating agent with more than 60 curated scientific skills, connectors, and databases spanning areas such as genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. It can work with Python and R, query connected research systems, and coordinate specialist agents for narrower tasks.

The compute model is unusually practical. Analyses can run locally, over SSH, or through an HPC environment, so the workbench can sit closer to the data and packages researchers already use. Anthropic also documents scientific renderers for protein structures, genomic tracks, chemical structures, and other domain artifacts instead of treating every result as plain text.

Reproducibility is a central part of the product pitch. Claude Science produces figures, tables, notebooks, and manuscripts alongside the code and environment history used to create them. A reviewer agent checks citations and calculations and can flag problems before the researcher accepts the result.

Strengths

The strongest idea here is integration. Scientific work often fragments across literature search, notebooks, specialist databases, cluster terminals, visualization tools, and manuscript editing. Claude Science gives that workflow one coordination layer while preserving inspectable artifacts and code.

It also has a better review shape than a generic research chatbot. Plans, provenance, editable figures, and a separate reviewer agent make it easier to inspect how a result was produced and where independent validation is still needed.

Limitations

Claude Science is still a beta, currently limited to macOS 13 or later and x64 Linux, and Team or Enterprise access requires admin enablement. The built-in ecosystem also leans heavily toward biology and biomedical research, so researchers in other fields should verify that the relevant packages, databases, and renderers are actually supported.

Local execution does not mean every part of the workflow stays on the device. Model context is still sent to Claude, and connected databases or compute services keep their own permission and data-handling boundaries. Reviewer agents can also be wrong. Citation checks and reproducible code help, but they do not replace domain-expert review, peer review, experimental validation, or clinical and regulatory controls.

Practical Tips

Start with a bounded, independently checkable workflow: reproduce a known analysis, regenerate an existing figure, or compare a literature summary against a trusted reference set. Confirm which data leaves the local or cluster environment before using sensitive material, and scope connectors to the smallest useful dataset.

Keep the generated notebook, environment details, source links, and review notes together. The product is most valuable when its provenance becomes part of the lab’s normal validation process rather than a decorative audit trail nobody reads.

Verdict

Claude Science is a serious early attempt to make an agent useful across the full scientific workflow rather than only at the literature-summary stage. It is most compelling for computational and life-science teams that already use notebooks, specialist databases, and shared compute—and that will still validate every important result independently.