Anthropic launches Claude Science as an auditable research workbench

Anthropic launched Claude Science in beta on June 30 for macOS and Linux. The announcement describes a research workbench that combines literature analysis, code execution, figures, manuscripts, scientific connectors, reusable skills, and access to local or remote compute. It was offered to Pro, Max, Team, and Enterprise users, with admin enablement required for organizational plans.

The product treats provenance as part of the interface

Anthropic says Claude Science keeps an auditable history connecting outputs to the code and steps that produced them. A coordinating agent can use curated scientific skills and connectors, invoke specialist agents, and ask a reviewer agent to inspect citations, calculations, and figure-code consistency. Sessions can run on a laptop, Linux host, remote machine over SSH, or an HPC login node.

Those features target real reproducibility problems, but an agent-generated history is not automatically a reproducible experiment. Teams still need immutable input identifiers, environment manifests, random seeds, tool versions, raw outputs, and an independently rerunnable entry point. A reviewer agent can flag issues; it cannot replace domain review or validate an instrument it cannot observe.

Compute approval deserves its own control plane

The announcement says Claude Science asks before reaching new compute resources and lets users review or revoke decisions before jobs are submitted. Research administrators should verify how credentials are stored, which commands can run over SSH, where intermediate data travels, and how cluster quotas and cancellations behave.

Start with a completed pipeline whose expected artifacts are known. Run it through Claude Science, export the history, and reproduce the result in a clean environment without the live session. That exercise will show whether the workbench’s audit trail is a useful scientific artifact or only a readable account of an otherwise ephemeral run.