AI has the potential to dramatically accelerate the pace of scientific discovery and development of medical interventions. Since we began our life sciences journey last fall, we have worked to improve the capabilities of our models, create connections with the scientific ecosystem through MCPs and skills, and initiate partnerships to realize this potential.
Today we’re introducing the most significant extension of these efforts: Claude Science, an AI workbench for scientists. Claude Science is an app that integrates the tools and packages most commonly used by researchers, produces auditable artifacts, and provides flexible access to computing resources.
Introduction to Claude Science
Scientific research is often boring. Researchers must work across dozens of databases, each with its own schema, deal with file formats that require bespoke data pipelines and viewers, and migrate between tools such as PubMed, Jupyter, R, and cluster terminals.
Claude Science integrates these fragmented tools into a single research environment, allowing scientists to conduct all stages of their research. It helps you analyze the literature to perform multi-stage research, generate detailed artifacts, and iteratively refine figures and manuscripts until they are ready for publication. All output includes an auditable history of how it was created, so results can be verified and reproduced. Similar to Jupyter Notebooks, you can access Claude Science from wherever you’re already working, including locally on macOS or Linux, or on a remote machine via SSH or using an HPC login node.
Users interact with a generalist coordinator agent that has access to over 60 pre-configured, curated skills and connectors for genomics, single cells, proteomics, structural biology, chemoinformatics, and more. These agents can launch other agents and work with professional agents created by users. A reviewer agent then checks the citations and calculations and flags and corrects any errors.
Today, we’re releasing a beta version of Claude Science for Claude Pro, Max, Team, and Enterprise users, and we’ll continue to improve the platform as we gather user feedback.
structure
A wealth of scientific results that are fully reproducible. Because scientific research is inherently visual, Claude Science produces figures and manuscripts along with the code that created them. Natively render rich scientific artifacts including 3D protein structures, genome browser tracks, chemical structures, and more. You can chat with your agent about details and annotate figures and manuscripts inline, so they know what to address to get them ready for publication.
When you generate a diagram, Claude Science includes the exact code and environment that created the diagram, a plain explanation of how it was created, and a complete message history. This helps you understand your input and makes it easier to verify and reproduce your work even months later. When you ask Claude Science to edit a diagram in plain language (for example, remove gridlines or change the axes to a logarithmic scale), the agent will edit your own code.
Manage your compute and scale on demand. For example, for large-scale analyses, such as protein folding or running genomics pipelines on large datasets, researchers often need to shift their focus to setting up a compute job, waiting for the job to be submitted to a cluster, checking whether it succeeded or failed, and retrieving the results. Claude Science handles this process for you. You can create plans, ask questions before reaching new resources, confirm or cancel decisions before creating and submitting jobs to compute resources your lab is already using (your own HPC cluster over SSH, or a modal account for on-demand computing), and scale your analysis from a single GPU to hundreds as needed.
The agent operates within a running session that keeps the context in memory, so even large datasets only need to be loaded once. Because it runs on the lab’s own infrastructure (laptops, Linux boxes, HPC login nodes), there is no need to leave systems where large or sensitive datasets already reside, and only the context needed for each step of the analysis is sent to Claude. As the pipeline runs, review agents inspect the output, flagging incorrect citations, untraceable numbers, and numbers that don’t match the underlying code, and self-correcting as they go. You can fork the session at any time to compare the two approaches without losing the original thread.
Domain ready from day one. Scientific knowledge is scattered across hundreds of specialized sources. For example, in biology, related data may exist across resources such as UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO, each with its own schema and query language. Additionally, they exist in journals, preprint servers, and domain-specific open models. When you ask Claude Science in plain language, our expert agents query and synthesize all of these sources so you don’t have to navigate through them individually. Claude Science uses the skills of NVIDIA’s BioNeMo agent toolkit to natively connect to BioNeMo’s life science models and libraries, including Evo 2, Boltz-2, and OpenFold3.
Scientists already have reliable models, datasets, and pipelines. Claude Science can also connect to these, allowing you to save pipelines as reusable skills and use connectors to access your lab’s preferred tools, which will be inherited automatically in future sessions. This customization feature gives you access to Claude, your own data, and the validated tools you already use in one conversation. Claude Science benefits from the expertise and platforms of our partners, and more scientists are accessing our tools through Claude.
What scientists are working on in Claude Science
Over the past few months, researchers have been collaborating with Claude Science in beta on tasks including single-cell RNA sequencing, CRISPR screen design, protein structure prediction, and chemoinformatics.
Manifold Bio designs tissue-targeted drugs that are based in specific organs or cell types, ensuring that the drug works where it is needed and leaves the rest of the body unaffected. We then test how millions of candidate binders, corresponding to hundreds of targets, are distributed throughout the organism at once. Manifold used Claude Science to designate targets for his latest experiment. Claude Sciences evaluated each tissue and target for surface expression, transport, and safety, and ranked candidates against criteria learned by Manifold from its own internal data. What differentiates Claude Science from typical coding assistants is that it can do this end-to-end, collecting the right data and applying the right judgment using the context of past programs it incorporates, Manifold said.
Jérôme Lecoq, a neuroscientist at the Allen Institute, used Claude Science to build a multi-agent “computational review template” consisting of about 20 custom skills specifically for writing long-form reviews. Subagents read thousands of papers, extract central claims and important quantitative findings, and store them in a state of evidence database. The pipeline then builds the narrative arc, writing reviews section by section and delegating each to its own specialized sub-agent. Within each section, dedicated agents generate quantitative cross-study figures directly from the evidence database. A key component of the workflow enabled by Claude Science is the use of actor-critic pairs. One agent creates the content, and another reviewing agent evaluates the content for accuracy and citation fidelity.
Before Claude Science, it could take Lecoq’s team two years to write a review like this. He currently has about 10 reviews with citations checked by reviewer agents, many of them over 100 pages. The team is currently working with subject matter experts to further refine the AI-based critic agent.
Stephen Francis, an epidemiologist and associate professor at the Brain Tumor Center at the University of California, San Francisco, also used Claude Science to support research into the molecular epidemiology of gliomas, a type of primary tumor that develops from glial cells in the brain. His lab studies the genetic basis of how thousands of low-impact germline variants combine to form an individual’s susceptibility. Although this work predates Claude Science, Francis said the app has dramatically accelerated analysis, allowing for comprehensive germline scrutiny across multiple approaches in about a tenth of the time it took before. His group independently validated Claude Science’s results and confirmed that it can produce both rapid and robust analyses.
Start Claude Science
The Claude Science app is available in beta on macOS and Linux on Pro, Max, Team, and Enterprise plans. We share it early so scientists can start using it for real problems and tell us how to improve it.
Team and Enterprise users must have Claude Science enabled by their administrator. We currently have team plans that offer discounted seats in active scientific labs at academic and non-profit research institutions. Click here for more information.
We also support up to 50 Claude Science AI for Science projects and provide up to $30,000 in credits. Modal offers up to $2,000 of compute for select projects. We are looking for projects that cross disciplines and explore the boundaries of science, with an initial focus on biological and biomedical research. Applications will be accepted until July 15, 2026, and award notifications will be sent by July 31. The project will run from September 1st to December 1st, 2026. Please apply here.
Join the AI for Science Discourse community to stay up to date with our products, provide feedback, and learn from other members of the Claude Science community.
Get started with Claude Science at claude.com/science.
