Source-led article
Anthropic Debuts Claude Science Beta for Reproducible AI-Powered Research
Anthropic has announced the beta release of Claude Science, an AI workbench aimed at revolutionizing scientific research by enhancing reproducibility and streamlining complex workflows. Launched on June 30, 2026, the application operates on Anthropic’s existing Claude models and is specifically designed for researchers grappling with diverse databases, notebooks, and cluster terminals in fields such as genomics, proteomics, single-cell analysis, structural biology, and cheminformatics.
The platform’s core innovation lies in its multi-agent architecture. A generalist coordinating agent receives plain-language requests and delegates tasks to specialized agents, including those created by users. A crucial reviewer agent operates continuously, flagging incorrect citations, untraceable numbers, and figures that do not align with their underlying code, initiating self-correction as needed.
Key Features for Scientific Advancement
Claude Science integrates commonly used tools and packages, enabling comprehensive literature analysis, multi-step research execution, and the generation of detailed, publication-ready artifacts. It natively renders 3D protein structures, genome browser tracks, and chemical structures. Every generated figure includes the exact code, environment, plain-language description, and full message history, ensuring auditable and reproducible results months after creation. Users can edit figures in plain language, with the agent dynamically updating its code. Sessions can be forked to compare different approaches without losing original work.
For large-scale analyses, Claude Science can draft and seek approval for plans before utilizing external computing resources. It supports deployment on local macOS or Linux machines, remote machines via SSH, HPC login nodes, and Modal, allowing analyses to scale from a single GPU to hundreds as required. The architecture ensures that large or sensitive datasets remain within the lab’s infrastructure, with only necessary context sent to Claude for processing.
Integration and Customization
The workbench comes pre-configured with over 60 curated skills and connectors, querying and synthesizing information from specialized biological databases like UniProt, PDB, Ensembl, Reactome, ClinVar, and ChEMBL, as well as journals and preprint servers. Claude Science also leverages skills from NVIDIA’s BioNeMo Agent Toolkit, providing native connections to genomics foundation models like Evo 2, biomolecular interaction prediction tools like Boltz-2, and protein structure prediction via OpenFold3.
Users can extend Claude Science through Model Context Protocol (MCP) connectors for lab tools and save existing pipelines as reusable skills. These connectors and skills persist across sessions, allowing researchers to integrate their validated tools and data while Claude orchestrates the workflow. Early beta users have successfully applied the platform to single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction, and cheminformatics.
Key facts
| Feature | Description |
|---|---|
| Release Date | June 30, 2026 (beta) |
| Core Technology | Multi-agent AI system running on existing Claude models |
| Target Audience | Scientists and researchers in genomics, proteomics, cheminformatics, etc. |
| Key Benefit | Enhanced reproducibility, streamlined workflows, auditable research outputs |
This development is significant for Coruja’s Indian audience, particularly those in the AI Tools and Tech sectors. The emphasis on reproducible research addresses a critical challenge in scientific discovery, which can accelerate advancements in Indian biotechnology and pharmaceutical R&D. By offering a platform that integrates diverse scientific tools and ensures verifiable results, Claude Science could empower Indian researchers to conduct more efficient and reliable studies, potentially leading to breakthroughs in healthcare, agriculture, and other vital areas that align with India’s scientific missions.