Claude Science: What Anthropic’s New AI Workbench Means for Life Science Teams

If your team works in genomics, drug discovery, or any data-heavy corner of the life sciences, you know the friction. Maybe your research is scattered across a dozen databases, a pile of scripts, and tools that barely talk to each other. Or maybe you’re a small company looking to see how AI can empower you. On June 30th, 2026, Anthropic announced Claude Science, an AI workbench built to pull that work into one place.

Edit: As of July 2nd, we have conducted an initial security review on Claude Science, and at this time we recommend that it not be enabled for broad employee use. This blog post has been updated to reflect these findings and includes updated advice for companies that are looking to roll out a pilot of Claude Science.

Anthropic posted a short demo of Claude Science, which you can watch below.

What Is Claude Science?

Claude Science is an AI workbench for scientific research. Rather than jumping between separate tools for literature review, data analysis, and writing, researchers work in one environment that connects those stages. It targets a problem most scientists know well. Research is spread across dozens of databases, each with its own schema, and stitching them together by hand burns time that should go to the science.

Claude Science launched as a beta for paid plans (Pro, Max, Team, and Enterprise) and is only available for macOS and Linux. You can download it from http://claude.com/science.

A few things stand out for lab and computational work:

• Auditable artifacts. Every output includes the code that produced it, a plain-language summary, and the full message history, so results are reproducible instead of a black box.
• Native scientific visuals. It renders 3D protein structures, genome browser tracks, and chemical structures right in the workbench, instead of sending you off to another program.
• A built-in reviewer. A reviewer agent checks the work and flags bad citations and figures it cannot trace to a source.

Built for the Way Scientists Actually Work

Claude Science is tuned for real research domains. It comes pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Additionally, it ships with more than 60 curated skills and connects to established databases, including UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO.

It also integrates NVIDIA’s BioNeMo Agent Toolkit and specialized models like Evo 2, Boltz-2, and OpenFold3. It handles compute resources for you, scaling from a single GPU to hundreds as a job needs. And it runs on the infrastructure you already have, whether that is a laptop, an HPC cluster over SSH, or on-demand cloud compute through Modal.

Early Results Are Promising

Here are a few exciting examples Anthropic shared:

• Manifold Bio used it to help nominate targets for tissue-targeting medicines.
• Jérôme Lecoq at the Allen Institute built multi-agent review templates that cut computational review writing from about two years down to months.
• Stephen Francis at UCSF reported completing comprehensive germline workups in roughly one-tenth the time.

Your Data Stays Where It Is

Claude Science is built so large or sensitive datasets never leave the systems where they already live. The tool runs where your data runs, and only the context each step needs is sent to Claude. This setup works well for teams handling patient data, proprietary sequences, or anything under compliance rules.

Why Claude Science Needs a Careful Look

The same capabilities that make Claude Science useful are what require a careful look. It runs on your machines, reads your files, executes code, and reaches out to databases and the web. Those powers are worth vetting before the tool accesses company or research data. An agent that can read a project folder can also read a file it was never meant to see.

July 2, 2026 Update: Review Findings

Since this post first published, our team completed an initial security review of Claude Science.

Our review examined how Claude Science handles credentials, permissions, and data boundaries before it goes near a regulated environment. It is the standard we apply to any new agent tool through our AI deployment service, where picking the right tool, setting policy, and securing it come before rollout. The tool includes real safeguards, but the residual risk is still material for teams handling sensitive or regulated data.

Claude Science is more than a standard desktop application. It functions as a local AI agent runtime that can run code on user devices, manage Python and R environments, use MCP tools and plugins, access local folders, store credentials, connect to external Anthropic and scientific data services, install packages, and update itself. The application includes meaningful security controls, but the residual risk remains material in biotech environments, where systems may hold sensitive research data, unpublished IP, partner-confidential information, regulated data, or PHI.

At this time, we recommend that it not be enabled for broad employee use.

What Admins Should Do

What Happens Next

We will keep testing Claude Science as it matures out of beta, and we will update this guidance as our recommendation changes.

Whether you are new to Claude or already using it across your team, we can help you choose the right plan, set up your team, and train your people to use AI effectively and securely. Tools like Claude Science pay off most for teams that adopt them deliberately.

Want to know the moment Claude Science gets our green light, along with the latest in tech and life science news? Sign up for our newsletter for more updates.