Claude Science turns AI research help into an auditable workbench

AI-for-science tools are moving beyond helpful text answers toward environments that can leave an audit trail. Claude Science is Anthropic’s beta workbench for researchers, built to combine literature review, code execution, figure generation, manuscript iteration, compute management, and validation inside one session. The product target is the friction scientists know well: PubMed in one tab, Jupyter or R in another, specialized databases elsewhere, and a cluster terminal waiting for jobs.

The concrete feature set is what makes the launch worth watching. Claude Science ships with more than 60 curated skills and connectors for genomics, single-cell analysis, proteomics, structural biology, cheminformatics, and related domains. A general coordinating agent can call specialist agents, and users can create their own. A reviewer agent checks citations and calculations, flags numbers that cannot be traced, and compares figures against the code and data that produced them.

The workbench also addresses compute, which is where many scientific workflows break out of ordinary notebooks. Anthropic says Claude Science runs locally on macOS or Linux, connects over SSH, works from an HPC login node, and can use Modal for on-demand compute. For large jobs, the system can draft a plan, ask before accessing new resources, and submit work to the lab infrastructure the researcher already uses. The company says large or sensitive datasets can remain on local or lab systems, with only the context needed for each step sent to Claude.

Early examples show the intended depth. Manifold Bio used Claude Science to help nominate tissue-targeting medicine candidates by assessing surface expression, trafficking, and safety criteria. Jérôme Lecoq at the Allen Institute built a multi-agent computational review template that reads thousands of papers, extracts central claims and quantitative findings, and uses reviewer agents to check accuracy and citation fidelity. Stephen Francis at the UCSF Brain Tumor Center used the platform around glioma molecular epidemiology work, with Anthropic reporting roughly a tenfold speedup for comprehensive germline analyses after independent validation by the lab.

Claude Science is available in beta on macOS and Linux for Pro, Max, Team, and Enterprise users, with admin enablement required for Team and Enterprise accounts. Anthropic is also offering support for up to 50 AI for Science projects, with as much as $30,000 in credits per project and up to $2,000 in Modal compute for selected projects. Applications are open through July 15, 2026, notifications are planned for July 31, and projects run from September 1 to December 1. The real test will be whether auditability holds under messy lab conditions, because scientific adoption depends less on fluent answers than on reproducible code, traceable claims, and error checks that researchers can trust.