If you're a researcher doing wet-lab work or data analysis, you probably bounce between PubMed, Jupyter, R, and cluster terminals while juggling databases in a dozen formats. Anthropic's newly launched Claude Science wants to fold those fragments into a single research environment — the official tagline is "scientific Claude Code." Meanwhile, a YC-backed team has shipped an open-source alternative, OpenScience, which supports Chinese models like DeepSeek and GLM and is completely free. This post covers both side by side.

What Is Claude Science

Claude Science is a customizable research workbench app that runs locally (macOS / Linux) or on remote machines (SSH / HPC login nodes). It's not a chat box but an environment that can carry a complete research workflow: analyzing literature, executing multi-step research, generating figures and manuscript drafts — every artifact with an auditable reproducibility history.

Officially it's positioned as an "AI workbench for scientists," currently in beta and open to Claude Pro / Max / Team / Enterprise users.

Three Core Capabilities

1. Rich Research Artifacts, Reproducible End to End

Science is inherently visual. Claude Science natively renders research artifacts — 3D protein structures, genome browser tracks, chemical structures — right next to the code. Every figure it generates comes with:

• The complete code and runtime environment that produced it
• A plain-language description of how it was generated
• The full message history

You can ask it to revise figures in natural language ("remove the gridlines," "switch the axis to a log scale"), and it will edit its own code and run it again. Months later, you can still trace where every number came from.

2. Managed Compute That Scales on Demand

Big jobs (folding proteins, running genomics pipelines) usually mean manually configuring a cluster, shipping tasks, waiting for results, and pulling them back. Claude Science takes over that layer:

• It drafts a plan first and asks for your approval before provisioning new resources
• Jobs go over SSH to your lab's existing HPC cluster, or GPUs are spun up on demand through a Modal account (scaling from a single card to hundreds)
• While jobs run, a reviewer agent checks the output, flagging mis-cited references, numbers that don't add up, and figures inconsistent with the code — then self-corrects

Because the agent runs in a persistent session, very large datasets only need to be loaded once, and sensitive data never has to leave your lab's own infrastructure.

3. Domain Connectors Out of the Box

Biology data is scattered across dozens of repositories — UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO — each with its own schema and query language. Claude Science ships with a universal orchestration agent that can access 60+ preconfigured research skills and connectors, covering genomics, single-cell, proteomics, structural biology, cheminformatics, and more. It also connects natively to models like Evo 2, Boltz-2, and OpenFold3 through the NVIDIA BioNeMo Agent Toolkit.

Real Cases: From Two Years to a Few Weeks

Jérôme Lecoq, a neuroscientist at the Allen Institute, used Claude Science to build a "computational review template" containing about 20 custom skills. Sub-agents read thousands of papers, extract core claims and key quantitative findings into an evidence library, then generate the review and quantitative cross-study figures section by section. A review like this used to take two years; he now has about 10 of them, most over 100 pages, with every citation checked by the reviewer agent.

Stephen Francis at UCSF's brain tumor center used it for glioma molecular epidemiology research, compressing the full germline variant analysis across multiple methods down to about one-tenth of the original time.

Access and Requirements

• Availability: beta on macOS and Linux; requires a Claude Pro / Max / Team / Enterprise plan
• Team / Enterprise: an admin needs to enable Claude Science in the backend
• Research discount: discounted Team seats for labs at academic institutions and nonprofit research organizations
• AI for Science program funding: up to 50 projects, each with up to $30,000 in credits, plus up to 2,000 in compute credits from Modal. Applications close July 15, 2026, with results announced by July 31 and the program running September 1 through December 1
• Entry point: claude.com/science

The Free Open-Source Alternative: OpenScience

If you'd rather not be tied to a Claude plan, or want to use Chinese models, a YC-backed AI research team has shipped an open-source alternative: OpenScience.

Its positioning matches Claude Science — an AI research workbench covering the full pipeline from literature search, hypothesis generation, and code experiments to paper writing — with two key differences:

• Completely free and open source, no plan requirements
• Swappable models, supporting Chinese models such as DeepSeek and GLM — use whichever you prefer
• Ships with 250+ research skill packs

For researchers in China, it's a way around subscription and compliance barriers. One caveat: an open-source project's stability, and its maturity around managed compute like HPC/Modal, will lag Anthropic's official version.

The One-Line Takeaway

Claude Science makes "a Claude Code for research" real — reproducibility, managed compute, domain connectors, and a reviewer agent that self-checks are what set it apart from general coding assistants. If you're not in the Claude ecosystem, OpenScience is currently the closest free substitute. Worth a try for researchers.