Claude Science Beta Runs Locally on Mac, Keeping Research Data Private
What You Need to Know
- Claude Science runs locally on user machines, keeping sensitive research data off external servers.
- App connects to 60+ biology databases including PubMed and UniProt simultaneously without window switching.
- Outputs include reproducible records of code and environment used for experiments and publications.
- Beta access limited to paid Claude tiers; free users excluded from current release.
Anthropic has released a beta desktop application called Claude Science, targeting researchers on macOS and Linux who need a dedicated environment for scientific work. The app sits alongside the existing Claude AI and Code tools on Mac, but its scope is narrower and more specialized from the start.
The most interesting detail is not the app itself but what it does with data. Claude Science runs locally on a user’s machine or existing lab hardware, which means sensitive health and research data never touches an external server. For anyone working under institutional data policies, that distinction matters more than any feature list.
What the app actually does
The tool connects to over 60 built-in biology databases simultaneously, including PubMed and UniProt, removing the need to switch between windows or manually pull from separate sources. When it generates outputs like 3D protein structures or manuscript drafts, it logs the exact code and environment used, creating a reproducible record from early experiment to final publication. Scaling compute is also handled automatically, moving from a single GPU to a larger cluster without manual configuration.
Access is currently limited to paid tiers. Researchers on a Pro, Max, Team, or Enterprise plan can download the installer directly from the Claude Science product page. Free users are not included in the beta.
Anthropic’s framing here is less about AI capability and more about workflow consolidation, positioning the app as a daily workbench rather than a chat interface. Whether that pitch lands will depend on how well the local processing holds up against the kinds of datasets research labs actually use, something a beta period is presumably designed to test.
0 Comments