Anthropic launches scientist AI workbench Claude Science

According to Beating Monitoring, Anthropic has launched a beta version of the Claude Science client for researchers. This system can run directly on a local machine, via SSH remote access, or on HPC (high-performance computing) nodes, integrating scattered scientific databases, Jupyter, the R language, and computing terminals into a single environment.

Its biggest feature is that it addresses the “reproducibility” pain point in scientific research. Every chart, 3D protein, or molecular structure generated by the system comes with the exact code used to produce the results, the runtime environment, and the complete conversation history. Researchers only need to describe what they want in natural language to directly modify the charts, and they can trace back to verify every data point.

When it comes to compute scheduling, it can automatically submit and manage computing tasks to the lab’s HPC cluster or Modal cloud compute resources, supporting the automatic scaling of analysis tasks from a single GPU to hundreds. The system comes with built-in coordinator agents and reviewer agents. When the coordinator agent calls NVIDIA’s BioNeMo toolkit and more than 60 pre-configured scientific skills for analyzing gene and protein data, the independent reviewer agent is responsible for cross-checking the computation results and paper citations, enabling automatic error correction.

In real-world tests, the UCSF Brain Tumor Center increased the analysis speed of glioma research to 10 times (cutting the time to 1/10 of the original). The Allen Institute used its multi-agent flow to build computational review templates, efficiently advancing what previously required writing a hundred-page academic review over two years in a very short time, with the reviewer agents completing fidelity checks of all citations.

Currently, the system is available to Claude Pro, Max, Team, and Enterprise subscription users. Anthropic has also set up a special fund for 50 research projects, providing up to $30,000 in project funding, and its partner Modal will also provide up to $2,000 in matching compute funding for selected projects.