Anthropic was released this week. claude science. This is an app for scientists and is available in beta. It runs on Anthropic’s existing Claude model, not the new model. This app is targeted at researchers who are familiar with databases, notebooks, and cluster devices. Perform a multi-step study and record how each result was obtained. The beta version is available for Pro, Max, Team, and Enterprise plans.
Claude Science builds on Anthropic’s life sciences efforts from last fall. That early work connected Claude to the scientific ecosystem through MCP and skills.
What is Claude Science?
Claude Science is an AI workbench for research. It integrates the tools and packages most commonly used by researchers. Analyze the literature, perform multi-stage research, and produce detailed deliverables. Figures and manuscripts can be improved until they are ready for publication.
You speak in plain language with one generalist coordinator agent. Its agents have access to over 60 carefully selected skills and connectors. These are preconfigured for genomics, single cell, proteomics, structural biology, and chemoinformatics.
Can be run locally on macOS or Linux. You can also work on remote machines via SSH or HPC login nodes. All output includes an auditable history of how it was created.
How multi-agent architecture works
A generalist coordination agent will receive your request in plain language. Other agents can be started to process work. You can also work with professional agents that you create yourself. NVIDIA describes these as preconfigured, domain-specific agents. Each knows the established workflow in their field.
A separate reviewer agent runs when the pipeline runs. Inspect the output step by step. Flag incorrect quotes and untraceable numbers. It also flags diagrams that don’t match the underlying code. After that, it will automatically correct itself as you progress.
Reproducibility and provenance
Scientific research is inherently visual. Claude Science therefore produces figures and manuscripts together with the code that created them. Natively render 3D protein structures, genome browser tracks, chemical structures, and more.
Generating a figure records your exact code and environment. It also records plain language explanations and complete message history. This makes it easier to verify and reproduce your work months later.
Edit diagrams in easy-to-understand language. For example, you can request that the axes change to logarithmic scale. The agent then edits its own code. You can also fork a session to compare two approaches without losing the original approach.
Compute that scales on demand
Large-scale analysis often requires more than a laptop. Protein folding is one example. Claude Science drafts a plan before reaching new resources. You can ask for approval and confirm or reverse your decision. Next, create and submit jobs to your own infrastructure.
This means HPC clusters or modal accounts over SSH. Analytics can scale from one GPU to hundreds of GPUs as needed. The agent keeps the context in memory, so large datasets are loaded only once.
The app runs on the lab’s own infrastructure. There’s no need to leave your very large or sensitive datasets off your current systems. Only the context needed for each step is sent to Claude.
Domain coverage and NVIDIA BioNeMo
Scientific knowledge is scattered across hundreds of specialized sources. In biology, this includes UniProt, PDB, Ensembl, and Reactome. This also includes ClinVar, ChEMBL, GEO, journals, and preprint servers. Specialist agents query and synthesize these sources.
Claude Science also uses skills from NVIDIA’s BioNeMo Agent Toolkit. This toolkit packages GPU acceleration functionality as a callable skill. It connects natively to Evo 2, Boltz-2, and OpenFold3. Evo 2 is a genomics-based model. Bolz-2 handles the prediction of biomolecular interactions. OpenFold3 handles protein structure prediction.
Usage and examples
Beta users performed single-cell RNA-seq analysis and CRISPR screen design. They also performed protein structure prediction and cheminformatics.
- Target nomination: Manifold Bio designs tissue-targeted medicines. Claude science was used to designate targets for the latest experiments. The app evaluated surface expression, transport, and safety for each tissue and target. We then ranked the candidates against Manifold’s own criteria. Manifold said the app did this end-to-end, unlike typical coding assistants.
- Long literature review: Jérôme Lecoq of the Allen Institute built a computational review template. This consisted of about 20 custom skills for long-form reviews. The subagent reads thousands of papers and stores them in the evidence state database. The pipeline then used a pair of actor and critic agents to write each section. Such a review once took his team two years. He currently has about 10 reviews, many of them over 100 pages.
- genomic epidemiology: UCSF’s Stephen Francis studies the molecular epidemiology of gliomas. Claude Science now performs germline testing in about a tenth of the time it took before. His group independently verified the results.
Comparison table
| size | claude science | General AI assistant | claude code |
|---|---|---|---|
| Main uses | Scientific research workflow | Q&A and drafting | software development |
| Run the actual pipeline | Yes, end to end | no | Yes, it’s code centric |
| Access to scientific databases | 60+ databases and skills | no | no |
| computing management | local, HPC (SSH), modal | no | local terminal |
| Reproducibility/source | Complete record per artifact | no | History of Git |
| Check citations and numbers | reviewer agent | no | no |
| native science renderer | protein, track, molecule | no | no |
| underlying model | Existing Claude model | Existing Claude model | Existing Claude model |
Expanding Claude Science
Claude Science is an app, so there is no separate inference API. Extend through connectors and skills that persist between sessions.
Lab tools connect through Model Context Protocol (MCP) connectors. This is the standard MCP client configuration format.
{
"mcpServers": {
"lab-eln": {
"command": "npx",
"args": ["-y", "@lab/eln-mcp-server"],
"env": { "ELN_API_KEY": "REPLACE_ME" }
}
}
}Save your existing pipeline as a reusable skill. What are skills? SKILL.md file:
---
name: rnaseq-qc
description: Run the lab's standard RNA-seq quality-control pipeline on a FASTQ directory.
---
# RNA-seq QC
1. Run `pipelines/qc.sh `.
2. Summarize the per-sample metrics.
3. Flag any sample below the QC threshold. Future sessions will automatically inherit these connectors and skills. So you keep your validated tools and data while Claude adjusts them.
Important points
- Claude Science is a beta app for macOS and Linux. Works on Anthropic’s existing Claude model.
- A coordinating agent delegates the work, and another reviewing agent checks the quotes, figures, and figures.
- All diagrams come with exact code, environment, description, and complete message history.
- Compute runs locally, HPC over SSH, or modally, and can scale from one GPU to hundreds of GPUs.
- Comes with 60+ databases and NVIDIA BioNeMo skills (Evo 2, Boltz-2, OpenFold3) for life sciences.
Please check Click here for technical details. Please feel free to follow us too Twitter Don’t forget to join us 150k+ML subreddit and subscribe our newsletter. hang on! Are you on telegram? You can now also participate by telegram.
Need to partner with us to promote your GitHub repository, Hug Face Page, product release, webinar, etc.? connect with us
Michal Sutter is a data science expert with a master’s degree in data science from the University of Padova. With a strong foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
