Claude Science: an AI workbench for scientists | Keryc
From making academic paperwork friendlier to speeding up analyses that used to take months, Claude Science arrives as an AI workspace built for scientific research. Think of a single place where tools, data and compute come together so you can focus on the science, not window juggling.
What is Claude Science
Claude Science is an app that brings together, in one place, the tools and packages scientists use every day. Think PubMed, Jupyter, R, 3D structure viewers, and cluster terminals all working inside the same session. The goal? So you can analyze literature, run complex pipelines, and polish figures and manuscripts without jumping between 20 different windows.
The platform produces rich, auditable scientific artifacts: 3D figures, genomic tracks, chemical structures, and the exact code that created them. Everything carries a history of how it was produced, so you can validate and reproduce results even months later.
Each output comes with the code, the environment, and a plain-language description of how it was obtained.
How it helps in your daily work
General and specialized agents: you interact with a coordinating agent that has access to more than 60 preconfigured skills and connectors for genomics, proteomics, structural biology, chemoinformatics, and more.
Automatic reviewer: a reviewer agent checks citations, calculations, and consistency between figures and code, and can fix errors on the fly. How many hours do you save by avoiding manual citation hunts?
Native reproducibility: when you ask for a change to a figure in natural language, the agent edits its own code and shows you exactly what changed.
Compute management and privacy
Claude Science handles the heavy lifting of compute resources: it plans, asks before provisioning new machines, and can submit jobs to your infrastructure (SSH to your HPC) or to cloud providers like Modal. It scales from a single GPU to hundreds as needed.
Important: it runs on your machines or on remote nodes you already use, so sensitive data doesn't have to leave your infrastructure. Only the minimal necessary context is shared between agents.
Ready for scientific domains from day one
The app connects to and queries specialized sources like UniProt, PDB, Ensembl, ClinVar, ChEMBL, GEO, and open-domain models. It also integrates NVIDIA BioNeMo tools to access models like Evo 2, Boltz-2 and OpenFold3.
If your lab already has validated pipelines or models, you can hook them up as reusable skills so future sessions inherit them automatically. In other words, Claude isn't here to replace your trusted resources; it's here to make them work together.
Real-world cases that show its potential
Manifold Bio used Claude Science to prioritize candidates in tissue-targeted drug discovery, combining public data with internal safety criteria.
At the Allen Institute, Jérôme Lecoq built a multi-agent chain to write long reviews: subagents extract claims and quantitative data from thousands of papers, and a reviewer agent verifies citations and figures. What used to take up to two years now moves much faster.
At UCSF, Stephen Francis sped up molecular epidemiology analyses, reducing stages that once took a long time to fractions of that time, with results independently validated by the lab.
Availability and how to get started
Claude Science launches today in beta for Claude Pro, Max, Team and Enterprise users on macOS and Linux. Teams should ask their admin to enable the app.
Anthropic also offers support for projects: up to 50 projects with credits of up to $30,000 each, and Modal contributes up to $2,000 in compute for selected projects. Application deadline: July 15, 2026; notifications on July 31. Selected projects will run from September 1 to December 1, 2026.
To stay updated, give feedback, and learn from other users, there's an AI for Science Discourse community and starter resources at claude.com/science.
Final reflection
Is this a replacement for scientists? No. It's a toolbox that reduces friction, automates verification, and lets teams focus on hypotheses and experimental design. If you're a researcher, professor, or part of a lab, the practical question is: what tedious tasks in your workflow could you delegate today so you can do science faster and with fewer errors?
