Anthropics launched the beta version of Claude Science, a new software interface designed for life science researchers. It does not rely on a newly trained model; instead, it directs existing Claude models to manage complex workflows in genomics, proteomics, and cheminformatics. The tool is currently available for Pro, Max, Team, and Enterprise subscribers.
What is Claude Science
The application acts as a central hub for the digital tools scientists rely on. It connects to over 60 pre-configured skills covering specific disciplines, allowing users to execute multi-step research tasks using plain English. The system generates detailed artifacts, including publication-ready figures and manuscripts, which users can refine through natural language instructions.
Researchers can operate the software locally on macOS or Linux systems. It also supports remote connections via SSH or high-performance computing (HPC) login nodes. Every result produced carries a complete, auditable history of the steps taken to generate it.
How the multi-agent system functions
A single coordinating agent receives the user’s request. This agent then activates specialist sub-agents to handle specific tasks. NVIDIA describes these specialists as preconfigured tools that understand established workflows within their specific fields.
A separate reviewer agent monitors the process as the pipeline runs. It checks outputs step by step, flagging untraceable numbers or incorrect citations. The agent also identifies figures that do not match their underlying code and attempts to self-correct errors before finalising the output.
Reproducibility and provenance
Scientific visualisation is central to the tool. Claude Science generates figures and manuscripts alongside the code used to create them. It renders 3D protein structures, genome browser tracks, and chemical structures natively.
When a figure is produced, the system records the exact code, environment, and a plain-language description. It also saves the full message history. This documentation makes it easier to validate and reproduce the work months later.
Users can edit figures by asking for changes, such as switching an axis to a log scale. The agent then modifies its own code to implement the request. Researchers can also fork a session to compare two different approaches without losing the original data.
Compute that Scales on Demand
Complex analyses often require more power than a standard laptop provides. Protein folding is one example where this is necessary. Claude Science creates a plan before requesting new resources. It asks for user approval and allows the researcher to review or revoke any decision before the job is submitted to their infrastructure.
This supports local HPC clusters over SSH or cloud accounts like Modal. The analysis can scale from a single GPU to hundreds as needed. Because agents maintain context in memory, large datasets load only once.
The application runs on the laboratory’s own infrastructure. Large or sensitive datasets never leave their current systems. Only the specific context required for each step is sent to Claude.
Domain coverage and NVIDIA BioNeMo
Scientific knowledge is fragmented across hundreds of sources. In biology, these include UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, journals, and preprint servers. Specialist agents query and synthesise information across these platforms for the user.
Claude Science also integrates skills from NVIDIA’s BioNeMo Agent Toolkit. This package offers GPU-accelerated capabilities as callable skills. It connects directly to Evo 2, Boltz-2, and OpenFold3. Evo 2 is a genomics foundation model. Boltz-2 predicts biomolecular interactions. OpenFold3 handles protein structure prediction.
Use cases with examples
Beta users have performed single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction, and cheminformatics tasks.
- Target nomination: Manifold Bio designs medicines that target specific tissues. It used Claude Science to identify targets for its latest experiments. For each tissue and target, the app assessed surface expression, trafficking, and safety. It then ranked candidates against Manifold’s proprietary criteria. The company noted the app performed this end to end, unlike general coding assistants.
- Long-form literature review: Jérôme Lecoq at the Allen Institute built a computational review template. It included about 20 custom skills for long-form reviews. Sub-agents read thousands of papers into an evidence state database. The pipeline then wrote each section using actor-critic agent pairs. Such reviews once took his team as long as two years. He now produces about 10 reviews, many over 100 pages.
- Genomic epidemiology: Stephen Francis at UCSF studies the molecular epidemiology of glioma. Claude Science ran germline workups in roughly one-tenth the previous time. His group independently validated the results.
Comparison of features
The following table contrasts the new tool with general assistants and coding-focused alternatives.
| 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
As an application, it does not have a separate inference API. Users extend it through connectors and skills, which persist across sessions.
Researchers can connect a lab tool via a Model Context Protocol (MCP) connector. This uses the standard MCP client config format:
{
"mcpServers": {
"lab-eln": {
"command": "npx",
"args": ["-y", "@lab/eln-mcp-server"],
"env": { "ELN_API_KEY": "REPLACE_ME" }
}
}
}Users can save existing pipelines as reusable skills. A skill consists of a folder containing a 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 <fastq_dir>`. 2. Summarize the per-sample metrics. 3. Flag any sample below the QC threshold.
Future sessions inherit these connectors and skills automatically. This allows researchers to keep their validated tools and data while Claude orchestrates them.
What it means
This release shifts the burden of reproducibility from the researcher to the software. By automating the checking of citations, numbers, and figure consistency, it reduces the manual verification required in modern bioinformatics. The ability to fork sessions and edit generated figures directly removes the need to manually rewrite code to adjust visual outputs. Ultimately, it aims to turn complex, multi-stage research pipelines into tasks manageable through natural language conversation.
