Scientific research workflows have long been fragmented across databases, Jupyter notebooks, and cluster terminals. This week Anthropic released Claude Science, a beta application designed specifically for researchers, to address that fragmentation head-on. Built on the company’s existing Claude models rather than a new foundational model, Claude Science functions as an AI workbench that integrates the tools and packages life scientists use most — analyzing literature, executing multi-step analysis pipelines, and producing publication-ready figures and manuscripts. The beta is available now to users on Pro, Max, Team, and Enterprise plans.
What Is Claude Science? An AI Workbench for Researchers
Claude Science is an AI workbench purpose-built for research. It integrates frequently used tools and packages, supports analysis of scientific literature, executes multi-step research, and produces detailed artifacts. Users can refine figures and manuscripts interactively until they are ready for publication. A single generalist coordinating agent accepts plain-language requests, then delegates tasks across more than 60 curated skills and connectors that come pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. The application runs locally on macOS or Linux, and can also connect to remote machines over SSH or to an HPC login node. Every output carries an auditable history of exactly how it was produced.
How the Multi-Agent Architecture Works
The coordinating agent receives a plain-language request and can spin up additional agents to handle specific tasks, including specialist agents that users can create themselves. NVIDIA describes these as preconfigured, domain-specialized agents, each knowing the established workflows for its field. A separate reviewer agent runs as the pipeline executes, inspecting every output step by step. It flags incorrect citations, numbers it cannot trace, and figures that do not match their underlying code, then self-corrects as it proceeds. This layered agent architecture is designed to maintain rigor throughout the research process.
Reproducibility and Provenance as Core Features
Scientific research is inherently visual, and Claude Science generates figures and manuscripts alongside the code that created them. It natively renders 3D protein structures, genome browser tracks, and chemical structures. When it generates a figure, it records the exact code, the environment, a plain-language description, and the full message history. This makes validation and reproduction straightforward months later. Users can edit figures in plain language — for example, asking the agent to change an axis to log scale — and the agent rewrites its own code. Sessions can be forked to compare approaches without losing the original work.
Compute That Scales on Demand While Keeping Data Private
Large analyses such as protein folding require more compute than a laptop can provide. Claude Science drafts a plan before reaching out to new resources, asks for approval, and allows the user to review or revoke any decision. It then writes and submits the job to the user’s own infrastructure — an HPC cluster over SSH or a Modal account. The analysis scales from a single GPU to hundreds as needed. Because agents hold context in memory, a large dataset loads only once. The application runs on the lab’s own infrastructure, so large or sensitive datasets never leave their current systems. Only the context needed for each step is sent to Claude.
Domain Coverage and Integration with NVIDIA BioNeMo
Scientific knowledge is distributed across hundreds of specialized sources. In biology, these include UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, journals, and preprint servers. Specialist agents query and synthesize across these sources. Claude Science also uses skills from NVIDIA’s BioNeMo Agent Toolkit, which packages GPU-accelerated capabilities as callable skills. This provides native connectivity to Evo 2 (a genomics foundation model), Boltz-2 (biomolecular interaction prediction), and OpenFold3 (protein structure prediction).
Real-World Use Cases from Beta Users
Beta users have already run single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction, and cheminformatics workflows. Manifold Bio, a company designing tissue-targeting medicines, used Claude Science to nominate targets for its latest experiments. For each tissue and target, the application assessed surface expression, trafficking, and safety, then ranked candidates against Manifold’s proprietary criteria — end to end, unlike a general coding assistant. Jérôme Lecoq at the Allen Institute built a computational review template comprising about 20 custom skills for long-form literature reviews. Sub-agents read thousands of papers into an evidence state database, then a pipeline wrote each section using actor-critic agent pairs. Such reviews once took his team up to two years; he now has about ten reviews, many exceeding 100 pages. Stephen Francis at UCSF used Claude Science to study the molecular epidemiology of glioma, running germline workups in roughly one-tenth the prior time, with his group independently validating the results.
Claude Science vs General AI Assistants vs Claude Code
| Dimension | Claude Science | General AI Assistant | Claude Code |
|---|---|---|---|
| Primary use | Scientific research workflows | Q&A and drafting | Software development |
| Runs real pipelines | Yes, end to end | No | Yes, code-focused |
| Scientific database access | 60+ databases and skills | No | No |
| Compute management | Local, HPC (SSH), Modal | No | Local terminal |
| Reproducibility / provenance | Full record per artifact | No | Git history |
| Citation and number checking | Reviewer agent | No | No |
| Native scientific renderers | Proteins, tracks, molecules | No | No |
| Underlying model | Existing Claude models | Existing Claude models | Existing Claude models |
Extending Claude Science Through Connectors and Skills
Claude Science is an application and does not expose a separate inference API. Instead, users extend it through connectors and skills that persist across sessions. A lab tool can be connected via a Model Context Protocol (MCP) connector using the standard client configuration format. Existing pipelines can be saved as reusable skills — a folder containing a SKILL.md file that describes the workflow. Future sessions inherit these connectors and skills automatically, allowing validated tools and data to persist while Claude orchestrates them.
What This Means for Researchers Now
Claude Science is a beta application for macOS and Linux running on Anthropic’s existing Claude models. A coordinating agent delegates work while a separate reviewer agent checks citations, numbers, and figures. Every figure ships with its exact code, environment, description, and full message history. Compute runs locally, on HPC over SSH, or on Modal, scaling from one GPU to hundreds. The application ships with more than 60 databases and NVIDIA BioNeMo skills including Evo 2, Boltz-2, and OpenFold3 for life sciences. Researchers who want to try it now can sign up through Anthropic’s Pro, Max, Team, or Enterprise plans. The technical details are available in Anthropic’s documentation, and the beta is actively accepting users working on macOS or Linux with access to their own infrastructure.
