Anthropic launched Claude Science on Tuesday, an application that puts protein-structure viewers, genome browser tracks, and chemical-structure rendering inside one working session instead of a dozen separate programs. The app is in beta for Claude Pro, Max, Team, and Enterprise subscribers on macOS and Linux, and it runs locally, over SSH to a lab cluster, or on an HPC login node.
The pitch is consolidation. A working scientist today moves between PubMed for literature, Jupyter or R for analysis, a cluster terminal for compute jobs, and a rotating set of viewers for whatever file format the data happens to use. Anthropic is betting that folding all of it into one agent-driven environment, rather than shipping another point tool, is what actually saves researchers time.
Claude Science routes requests through a coordinating agent with access to more than 60 pre-configured skills and connectors covering genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. That agent can spawn specialist sub-agents, and a separate reviewer agent checks the outputs, flagging incorrect citations or numbers that do not trace back to the underlying code.
Every figure the app produces ships with the code, environment, and full message history that generated it, which Anthropic says makes results reproducible months later without a researcher having to reconstruct the pipeline from memory. Users can request plain-language edits, such as switching an axis to log scale, and the agent rewrites its own code rather than requiring a manual patch.
Compute handling is the second pillar. Large jobs, protein folding or a genomics run across a large dataset, typically force a researcher to manually configure a cluster job, wait in a queue, then pull results back by hand. Claude Science drafts the compute plan, asks permission before touching new resources, and submits the job to whatever infrastructure the lab already has, a private HPC cluster over SSH or Modal for on-demand GPUs. Data stays on the researcher’s own systems; only the context each step needs is sent to Claude.
The domain integration draws on outside partners. Anthropic said the app connects natively to Nvidia’s BioNeMo Agent Toolkit, giving it access to life-science models including Evo 2, Boltz-2, and OpenFold3, on top of database connectors for UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO.
Anthropic’s own case studies describe substantial time savings. The company said Jérôme Lecoq, a neuroscientist at the Allen Institute, built a 20-skill multi-agent pipeline in Claude Science that writes literature reviews section by section, with actor-critic agent pairs checking citation accuracy. Lecoq’s team has produced roughly ten reviews, some over 100 pages, in what Anthropic said previously took as long as two years per review. Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, told Anthropic the app cut germline variant analysis time for glioma research to roughly one-tenth of its prior duration, with his lab independently validating the results.
Those figures come from Anthropic’s own announcement and from labs the company chose to highlight, not from an independent audit of adoption or accuracy across its broader beta pool. The company has not published failure rates for its reviewer agent or disclosed how many beta users have adopted the tool beyond the three labs named in the post.
Anthropic is also opening applications for up to 50 “AI for Science” projects, offering as much as $30,000 in credits plus $2,000 in Modal compute per selected project, with a July 15 application deadline and awards announced by July 31. Selected projects run from September through December.
For research teams currently stitching together Jupyter, cluster access, and separate structural-biology viewers, Claude Science is worth a pilot on a real dataset before the beta closes, since the reproducibility claims and the reviewer agent’s error-catching are the parts most likely to determine whether it replaces existing lab workflows or just adds another tool to the pile.
Reported by Anthropic on July 1, 2026.
