You've probably already used Claude Code: you hand it a task in natural language, and it writes and then runs the code for you. Claude Science rests on the same principle, but applied to a researcher's entire job: querying dozens of scientific databases, launching heavy computations, generating figures and papers, then checking every citation. Where Claude Code helps you ship code solo, Claude Science helps a scientist run their whole research workflow end-to-end.
Claude Science is therefore an AI-powered workbench built by Anthropic to accelerate scientific research. It brings together in a single environment the tools, databases and computing resources researchers use day to day, while producing traceable and reproducible results.
Available in beta since June 30, 2026 for Pro, Max, Team and Enterprise users, it aims to turn often-fragmented work into a unified research workflow.
In this article, we'll explore in detail what Claude Science is, how it works, who it's for, real and concrete use cases, and how to get started today.
You don't need to be a researcher or a doctor to grasp the idea: we explain every technical term as we go. Focus on the principle, the vocabulary will follow.
What is Claude Science?
Claude Science is an application that integrates the most widely used tools and libraries for researchers within a single research environment.
Instead of juggling dozens of databases, file formats and different tools (PubMed, Jupyter, R, cluster terminals, etc.), scientists can carry out every stage of their work there.
The starting point is simple: scientific research often involves tedious and repetitive tasks. Each database has its own schema, each file format requires bespoke data pipelines, and switching from one tool to another wastes considerable time. Claude Science sets out to remove this friction.
Claude Science runs where you already work: locally on macOS or Linux, or on a remote machine via SSH or an HPC login node (high-performance computing).
The app relies on a generalist coordinating agent with access to more than 60 skills and connectors preconfigured for genomics, single-cell analysis (at the scale of the individual cell), proteomics, structural biology, cheminformatics, and many other domains. These agents can create others and collaborate with specialist agents built by users.
The role of the reviewer agent
What sets Claude Science apart in particular is its reviewer agent. It checks citations and calculations, flags errors and corrects them automatically as the analysis unfolds.
This addresses a major challenge of scientific reliability:
- limiting citation errors,
- untraceable numbers,
- and inconsistencies between a figure and the code that produced it.
And in practice? In research, a wrong citation or a number nobody can trace back can invalidate months of work, or even get an already-published paper retracted. An agent that monitors all of this continuously means time and credibility saved.
How does Claude Science work?
Claude Science is built around three pillars: the production of rich, reproducible scientific artifacts (figures, manuscripts, code...), intelligent compute management, and ready-to-use, domain-specific preconfiguration.
Rich, reproducible scientific artifacts
Scientific research is visual by nature. Claude Science therefore generates figures and manuscripts together with the code that created them.
It natively renders rich artifacts: 3D protein structures, genome browser tracks, chemical structures, and more.
Each generated figure comes with three essential elements:
- The exact code and environment that produced it
- A plain-language description of how it was created
- The complete message history
This traceability lets you understand the inputs and reproduce the work, even months later. You can also request changes in natural language, for example:
"Remove the gridlines from this chart"
"Change the Y axis to a log scale"The agent will then edit its own code to apply the changes.
On-demand compute management
Large-scale analyses (such as protein folding, meaning predicting the 3D shape a protein adopts, or running a genomics pipeline over a massive dataset) often force researchers to set up a compute job, wait for it to be sent to a cluster, check whether it succeeded or failed, then pull back the results. Claude Science automates this entire process.
It drafts a plan, asks for your approval before accessing new resources, and lets you review or revoke any decision before submitting the job to your lab's computing resources (your HPC cluster via SSH, or your Modal account for on-demand compute). The analysis can thus scale from a single GPU to hundreds, as needed.
Because the agents work within an active session that keeps the context in memory, even massive datasets only need to be loaded once.
A point that's crucial for privacy: the app runs on your lab's infrastructure (laptop, Linux machine or HPC login node), so large or sensitive datasets never need to leave the systems they already live on. Only the context needed for each step of the analysis is sent to Claude.
You can also fork (duplicate) a session at any time to compare two approaches without losing the original thread.
Ready for your field on day one
Scientific knowledge is scattered across hundreds of specialized sources.
In biology, for example, the relevant data may live in resources such as UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL or GEO (each with its own schema and query language) as well as in journals, preprint servers and specialized open models.
When you ask a question in natural language, specialist agents query and synthesize all of these sources for you. Claude Science notably draws on the skills of NVIDIA's BioNeMo Agent Toolkit to connect natively to life-sciences models and libraries, including Evo 2, Boltz-2 and OpenFold3.
In plain terms: Evo 2 is a DNA (genomics) model that can predict the effect of mutations, while Boltz-2 and OpenFold3 predict the 3D structure of proteins and their assemblies. Why does this matter? Predicting the effect of a mutation helps determine whether it can trigger a disease without having to test it in the lab, and knowing a protein's 3D shape is the first step toward designing a drug that can bind to it.
Scientists already have models, datasets and pipelines they trust. Claude Science can connect to them, save any pipeline as a reusable skill, or access your lab's preferred tool via a connector. Future sessions inherit them automatically.
| Pillar | Main benefit |
|---|---|
| Reproducible artifacts | Figures and manuscripts with code and full history |
| On-demand compute | Scalability from one GPU to hundreds, locally or via Modal |
| Domain preconfiguration | More than 60 scientific databases connected from the start |
| Reviewer agent | Automatic verification of citations and calculations |
Concrete use cases of Claude Science
Over the past few months, researchers have used Claude Science in beta for a variety of tasks: single-cell RNA sequencing analysis, CRISPR screen design (large-scale tests to pin down the role of many genes), protein structure prediction, cheminformatics, and much more.
Manifold Bio: drug targeting
Manifold Bio designs tissue-targeting drugs: they home in on a specific organ or cell type, so the drug acts only where it's needed and spares the rest of the body. The company tests how millions of candidates distribute through a living organism.
Manifold used Claude Science to identify the targets for its latest experiments. For each tissue and target, the tool assessed surface expression, intracellular trafficking and safety, ranking candidates against criteria drawn from its own proprietary data. What set Claude Science apart from a mere coding assistant was its ability to carry this work out end-to-end, gathering the right data and applying the right judgment with the context of past programs.
Allen Institute: computational reviews
Jérôme Lecoq, a neuroscientist at the Allen Institute, used Claude Science to build a multi-agent "computational review template" made up of around 20 custom skills, dedicated to writing long literature reviews (review articles).
The sub-agents read thousands of papers, extract the central claim and the key quantitative finding, then store them in an evidence database. The pipeline then builds a narrative arc, writing the review section by section, with each section delegated to a specialized sub-agent. A key mechanism is the use of actor-critic pairs: one agent creates the content while a separate reviewer agent evaluates its accuracy and citation fidelity.
Before Claude Science, writing such a review could take Lecoq's team up to two years. He now has about ten reviews, many of them over 100 pages, with citations checked by the reviewer agents.
UCSF: molecular epidemiology of glioma
Stephen Francis, an associate professor and epidemiologist at the UCSF Brain Tumor Center, used Claude Science to support his studies on the molecular epidemiology of glioma, a primary tumor that begins in the brain's glial cells. His lab studies how thousands of small-effect germline variants (small inherited genetic variations, each with a minimal impact) combine to shape individual susceptibility.
Although this work predates Claude Science, Francis says the app dramatically accelerated the analysis, making it possible to run complete germline workups in roughly one-tenth of the time previously required. His group independently validated the results, confirming that the tool produces analyses that are both fast and robust.
What Claude Science changes, even if you're not a researcher
Let's be honest: if you don't fold proteins on the weekend, you're not going to open Claude Science tomorrow morning. But this product says something that does concern you directly.
Until now, an AI assistant answered a question. Here, we cross a threshold: a generalist agent coordinates specialist agents, runs the computations itself on real infrastructure, checks its own work, and reuses the skills accumulated on previous projects. This is no longer a tool that assists, it's a tool that runs an entire workflow end-to-end.
And this is exactly the direction every AI tool is heading, not just in science. If this model works in the most demanding field there is (biomedical research, where a single citation error can derail everything), it will work on far simpler tasks: your marketing research, your content production, your project management.
For a solopreneur, the message is crystal clear. Yesterday, Claude Code already let you ship code without a dev team. Tomorrow, agents able to chain together complete workflows (research, computation, verification, production) will let a single person do the work of an entire team. Claude Science is a preview of that future, applied to the hardest terrain there is. It's up to you to imagine what it will look like on yours.
Best practices to get the most out of Claude Science
To make the most of this AI workbench, here are a few recommendations drawn from observed usage.
Always validate results
Even with a built-in reviewer agent, independent validation remains essential. As Stephen Francis's lab showed, confirming results with your own methods guarantees their robustness before publication.
Capitalize on reusable skills
Save your trusted pipelines as reusable skills. Your future sessions will inherit them automatically, which avoids reconfiguring the same tools for every project and speeds up your recurring analyses.
Use forking to compare approaches
When you're torn between two analytical methods, duplicate the session to explore them in parallel. That way you keep the original thread intact while testing alternatives.
Keep your sensitive data local
Take advantage of the fact that Claude Science runs on your own infrastructure. Confidential datasets never leave your systems: only the minimal context needed for each step is transmitted. This is a major asset for compliance and privacy in biomedical research.
How to get started with Claude Science?
The Claude Science app is available in beta on macOS and Linux for the Pro, Max, Team and Enterprise plans. Anthropic is sharing it early so that scientists can use it on real problems and send back their feedback.
- Team and Enterprise users: your administrator must enable Claude Science.
- Academic labs: a Team plan offers discounted seats for active scientific labs at academic institutions and nonprofit research organizations.
- AI for Science program: Anthropic supports up to 50 projects, providing up to $30,000 in credits, with Modal providing up to $2,000 in compute credits for selected projects.
The AI for Science program targets projects at the frontier of science, with an early focus on biology and biomedical research. Be sure to check the application and notification dates through Anthropic's official channels.
To stay up to date on product announcements, share your feedback and learn from the community, join Anthropic's AI for Science community space.
Frequently asked questions
Is Claude Science free?
Claude Science is available in beta for the paid Pro, Max, Team and Enterprise plans. It is not offered on the free plan. However, discounted seats exist for academic labs, and an AI for Science program offers compute credits to selected projects.
Which operating systems does Claude Science run on?
The app is available in beta on macOS and Linux. It can run locally or on a remote machine via SSH or an HPC login node, which makes it compatible with labs' high-performance computing infrastructures.
Is my sensitive data protected with Claude Science?
Yes. Because Claude Science runs on your own infrastructure (laptop, Linux machine or HPC login node), large or sensitive datasets never leave your systems. Only the minimal context needed for each step of the analysis is sent to Claude.
Can Claude Science connect to my existing tools?
Absolutely. You can connect your trusted models, datasets and pipelines, save any pipeline as a reusable skill, or access a tool via a connector. It supports more than 60 scientific databases and integrates with NVIDIA's BioNeMo Agent Toolkit (Evo 2, Boltz-2, OpenFold3).
How does Claude Science differ from a classic coding assistant?
Unlike a generalist coding assistant, Claude Science runs research end-to-end: it gathers the right data, applies sound judgment with the context of past work, manages compute, and produces traceable artifacts. Its reviewer agent continuously checks citations and calculations to guarantee scientific reliability.
How can I get trained on Claude Science?
We're preparing a course on this topic, but in the meantime, you can learn Claude Code with our course.