Anthropic Launches Claude Science, an AI Workbench Built for Scientists

Anthropic launched Claude Science on June 30. Available in beta today for Pro, Max, Team, and Enterprise subscribers, the AI workbench pulls PubMed, Jupyter, R, cluster terminals, and over 60 scientific databases into a single research environment on macOS and Linux. Every output carries an auditable trail of exactly how it was made.

One Environment for Dozens of Disconnected Research Tools

Context-switching kills research time. A biologist's workday might span PubMed for literature, a Jupyter notebook for analysis, an R script for statistics, and a terminal session on a cluster - none of which share context with each other. Claude Science collapses all of that into one place, with a generalist coordinating agentic AI at the center that hands tasks to specialist sub-agents and spins new ones as analysis demands shift.

60 pre-configured skills come out of the box. Connectors for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics are ready from day one, with NVIDIA's BioNeMo Agent Toolkit adding native access to Evo 2, Boltz-2, and OpenFold3. A reviewer agent runs in parallel across every session - catching citations that don't check out, flagging figures whose underlying code doesn't match what's displayed, and self-correcting as it goes rather than just leaving errors for human review.

Sensitive Data Stays in Your Lab, Claude Science Handles the Compute

Running a large genomics pipeline is a time sink. Setting up the job, submitting it to a cluster, monitoring whether it succeeded, and pulling results back can break a researcher's focus for hours - and none of that is the science. Claude Science handles the entire loop itself, submitting jobs to whatever infrastructure the lab already runs: a local machine, an HPC cluster over SSH, or Modal for on-demand GPU scaling from a single GPU to hundreds as the analysis demands.

Sensitive data stays put. Claude Science runs on your lab's own infrastructure, so datasets never leave the systems they already sit on - only the context for each step of the analysis travels to Claude. For labs working with IRB-protected patient data or proprietary compound libraries, that distinction matters considerably more than convenience. Researchers can also fork any session mid-run to compare two approaches without losing the original thread - useful when a hypothesis changes and you don't want to discard what came before.

How Manifold Bio, the Allen Institute, and UCSF Used Claude Science

Manifold Bio designs tissue-targeting medicines. Claude Science nominated targets for its latest experiments - assessing surface expression, trafficking, and safety for each candidate, then ranking them against criteria drawn from Manifold's own proprietary programs. Working end-to-end without handing off between tools was the distinction Manifold said a standard coding assistant couldn't match.

Two years per review. That was Jérôme Lecoq's timeline at the Allen Institute before Claude Science. He built a multi-agent review template from about 20 custom skills - sub-agents read thousands of papers, extracted the central claim and key finding from each, and stored them in an evidence database before specialist agents wrote the review section by section, generating cross-study figures directly from that database. Lecoq now has 10 reviews finished, many over 100 pages, with actor-critic agent pairs checking citation accuracy throughout: one agent creates content, a separate reviewer evaluates it for accuracy and fidelity.

Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, used Claude Science to study glioma - specifically how thousands of small-effect germline variants combine to shape individual cancer susceptibility. One-tenth the time. Francis's group completed comprehensive germline workups across multiple analysis approaches at that speed, then independently validated every result and confirmed no loss of rigor.

Beta Is Live Today, With $30,000 Grants Open Through July 15

Beta launches today on macOS and Linux. Pro, Max, Team, and Enterprise subscribers get access now, though Team and Enterprise accounts require admin activation. Anthropic is also offering discounted Team plan seats for active scientific labs at academic institutions and nonprofit research organizations.

First wave focuses on biology. Anthropic welcomes other domains but says biomedical research is the early priority for the initial cohort of 50 projects. MCP connectors let labs bring their own proprietary data and trusted pipelines into Claude Science, with future sessions inheriting those connections automatically - so configuring a lab's environment doesn't have to repeat each time. Grant applications close July 15.