Anthropic unveils 'Claude Science' for scientific research

A workbench, not just a chat window

Anthropic describes Claude Science as an app for scientific research that integrates commonly used tools and packages, produces auditable artifacts, and gives researchers flexible access to computing resources. The stated target is the fragmented workflow many labs know too well: PubMed in one tab, Jupyter in another, R somewhere else, a cluster terminal open, and a pile of file formats that each need their own gotcha-filled pipeline.

The product is available in beta for Claude Pro, Max, Team, and Enterprise users. Anthropic says it can be used locally on macOS or Linux, or on a remote machine over SSH or with an HPC login node. That matters because the hardest part of many research-assistant demos is not generating a paragraph — it is getting the assistant close enough to the actual execution environment without turning the pipeline into unreproducible glue.

Claude Science is built around a generalist coordinating agent with access to more than 60 curated skills and connectors. Anthropic says these are preconfigured for areas including genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. The agent can also spin up other agents and work with specialist agents created by users.

Reproducibility is the key engineering claim

The most important claim for our audience is the audit trail. Anthropic says every output carries a history of how it was made, so users can validate and reproduce results. When Claude Science generates a figure, it includes the code and environment that produced it, a plain-language description of the process, and the full message history.

That is the right abstraction to watch. In practical terms, scientific AI tooling fails when the nice final plot cannot be traced back to the exact input, command, package state, or notebook cell that created it. Anthropic is positioning Claude Science closer to a notebook-style research environment than an “ask-and-answer” assistant.

The app also renders scientific artifacts directly, including 3D protein structures, genome browser tracks, and chemical structures. Radical Data Science separately described the beta as a workbench for macOS and Linux with tools for rendering those artifact types. Anthropic says users can ask for figure edits in plain language — for example, removing gridlines or changing an axis to log scale — and the agent edits its own code.

There is also a reviewer agent. Anthropic says it checks citations and calculations, flags errors, and corrects them. Treat that as a useful sanity check layer, not a reason to skip review. The more automated the manuscript and figure loop becomes, the more we need to inspect the generated code path, not just the rendered output.

What to verify before putting it near real pipelines

For dataset and curation teams, the immediate evaluation should be boring and concrete. Can Claude Science preserve the provenance of a multi-step analysis across files, prompts, code, and environment state? Can it survive the unglamorous parts of real work: schema mismatches, failed jobs, missing viewers, and long-running compute?

Anthropic says the system can manage compute and scale on demand, including workflows such as folding a protein or running a genomics pipeline over a large dataset. It says Claude Science can handle the process of setting up a job, sending it to a cluster, checking success or failure, and pulling results back. That is exactly where we would test first, because cluster orchestration is full of edge cases that look simple in a demo and break in production.

The broader signal is also worth noting. A London School of Economics blog framed the need as dedicated research tools rather than “off the shelf” AI. Claude Science fits that direction: domain connectors, executable artifacts, scientific renderers, and review steps bundled into a research workbench.

Our practical take: evaluate it like infrastructure, not like a chatbot. Start with one reproducible pipeline, one known dataset, one expected figure, and one failed-job path. If the audit history, code, environment capture, and compute handoff remain inspectable after iteration, then Claude Science becomes interesting for serious research workflows. If those pieces blur, it is just more boilerplate wrapped around the same old reproducibility problem.