From chatty assistant to working lab console
Claude Science is an AI research workbench that pulls 60‑plus life sciences databases, coding tools, and compute resources into a single workspace where scientific AI agents can plan, launch, and audit complex experiments from start to finish. That shift matters: most AI research tools so far have helped scientists outline methods and draft papers, but left the real work of running pipelines, chasing citations, and wrestling clusters to humans. Anthropic’s new Claude Science workbench is a direct attack on that gap. By unifying databases, shells, notebooks, and remote high‑performance computing under one agentic interface, it turns the model from an overqualified note‑taker into an execution engine. The message is blunt: if your AI cannot submit the job and verify the citations, it’s not ready for the lab.
One AI console instead of a dozen disconnected tools
The most important change Claude Science brings is ruthless consolidation. Anthropic launched the app on June 30 as a single workspace that pulls a researcher’s scattered tools into one place and lets AI agents run the work end to end. Scientists today lose hours hopping between PubMed, Jupyter, R, and a cluster terminal; they fight custom pipelines for every file format. Claude Science tries to end this juggling by wiring more than 60 scientific databases and toolkits for genomics, single‑cell work, proteomics, structural biology, and cheminformatics into one environment. A coordinating agent sits in the middle, speaking plain language on one side and dispatching work to these specialized tools on the other. In design terms, this is not another app; it is an operating layer for experimental science, and that is exactly the layer AI should own.
Agentic experiments: from literature search to live compute jobs
Claude Science is opinionated about what AI should be doing in a lab: running experiments, not only brainstorming them. The coordinating agent hands tasks to specialist sub‑agents, while a separate reviewer agent checks every citation and calculation and fixes errors as it goes. In practice, that means a literature review that auto‑verifies every reference instead of sprinkling in hallucinated papers—a real response to the deluge of fabricated citations AI writing has pushed into drafts. Crucially, the workbench is wired into computation. It manages complex analyses and job submissions, talking to local machines, remote HPC clusters over SSH, and on‑demand platforms so large genomics pipelines or protein folding runs happen through the same interface. According to Anthropic, “It brings together databases, code, and computing power, then lets AI agents run the work from start to finish.”
Drug discovery automation and the life sciences stack
Claude Science is unapologetically built around life sciences first. It connects to more than 60 specialized tools and connectors for genomics, proteomics, cheminformatics, and structural biology, and it plugs into life sciences databases like UniProt, PDB, ChEMBL, and GEO through curated skills. For drug discovery automation, the crucial bridge is NVIDIA’s BioNeMo Agent Toolkit: Claude’s agents can call models such as Evo 2, Boltz‑2, and OpenFold3 to design, fold, or evaluate molecules, then feed those results straight into downstream analysis. In other words, Claude Science is the reasoning and orchestration layer, while specialized biology models stay where they were built. This is the right division of labor. Scientific AI agents should understand goals, design workflows, and maintain audit trails, not try to replace every simulation engine in computational biology.
Reproducibility, deployment flexibility, and what happens next
If AI is moving into the lab, it must leave a paper trail. Every artifact Claude Science produces is tied to its code, environment, plain‑language description, and full history so results can be traced and reproduced later. Reviewer agents catch citation errors and calculation issues along the way, turning reproducibility from an afterthought into a default. The deployment story is equally pragmatic: the beta runs on macOS and Linux, on local machines or remote systems via SSH, and it is available to Pro, Max, Team, and Enterprise users. Anthropic is backing its bet with compute: applications for dedicated AI for Science project credits are open until July 15, 2026, supporting up to 50 projects with substantial platform support. Early users report dramatic cuts in analysis time and better data validation, from CRISPR screen design to multi‑agent literature review. The direction is clear: the next wave of scientific progress will come from labs that treat AI agents as first‑class experimenters, not as glorified word processors.
