News · 2026-07-11

Anthropic launches Claude Science, an AI workbench that keeps data in the lab and checks its own citations

Anthropic launched Claude Science, an AI workbench aimed squarely at working researchers, built around two ideas that address the field's biggest hesitations about AI-in-the-lab: keep the data local, and make the AI check its own work. It integrates the tools scientists actually use -- PubMed, Jupyter, R, cluster terminals -- runs on the lab's own hardware so sensitive datasets never leave, and pairs a working agent with a separate reviewer agent that flags and corrects citation and calculation errors.

Key facts

• What: a beta workbench app for scientists on macOS and Linux, for Claude Pro, Max, Team, and Enterprise users.
• Scope: 60+ curated skills and connectors pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics.
• Trust mechanism: a reviewer agent that checks citations and calculations and self-corrects errors.
• Primary source: Anthropic's Claude Science announcement.

The design solves a specific tension. Scientists have real reasons not to pipe their work through a chatbot: proprietary or patient data can't leave the building, and an AI that confidently invents a citation is worse than no AI at all. Claude Science answers the first with local execution -- it runs on your laptop, Linux box, or HPC login node, and only sends Claude the minimal context each step needs, so large or sensitive datasets stay put. It answers the second with an actor-critic structure: a generalist coordinating agent does the work and dispatches specialist sub-agents, while a distinct reviewer agent audits the output for accuracy and citation fidelity. This is the same 'one agent produces, another checks' pattern that shows up across trustworthy-AI research, applied to the messy reality of a scientific workflow.

A few mechanics make it concrete. Every figure or manuscript it produces comes with the exact code and environment that generated it, a plain-language description, and the full message history -- so a result is reproducible, not a screenshot. You can edit figures in plain English ('change the axis to log scale') and it rewrites its own code. For heavy jobs it drafts a compute plan and asks before reaching new resources, letting you review or revoke before it submits work to your cluster over SSH or to Modal for on-demand GPUs, scaling from a single GPU to hundreds. It also connects natively to life-sciences models -- Evo 2, Boltz-2, OpenFold3 -- through NVIDIA's BioNeMo toolkit.

The researcher stories in Anthropic's announcement are the most persuasive part. Allen Institute neuroscientist Jerome Lecoq built a multi-agent 'computational review template' with about 20 custom skills: sub-agents read thousands of papers, extracted central claims and quantitative findings into an evidence database, and actor-critic pairs generated and then audited the writing. He says reviews that used to take up to two years now number roughly ten, many over 100 pages. UCSF epidemiologist Stephen Francis reports that analysis of glioma molecular epidemiology ran in roughly one-tenth the time -- and, tellingly, his group independently validated the results rather than taking them on faith. Manifold Bio used it to nominate drug targets end-to-end, ranking candidates against proprietary internal data.

Why it matters is that this is the AI-for-science pitch made practical rather than aspirational. Where SciReasoner makes a model's scientific reasoning inspectable, Claude Science makes the whole workflow auditable and keeps the human in the loop for compute and data. The honest caveat is that a reviewer agent is not a guarantee -- it catches errors, it does not certify correctness, and the researchers who trust it most are the ones, like Francis, who still independently validate. Anthropic is also seeding adoption directly: it committed to fund up to 50 projects with up to $30,000 in credits each, with applications open through July 15, 2026, and projects running September through December. That is a real forward commitment worth checking against later -- and a sign the company is betting the lab, not just the chat window, is where the next AI foothold is.

Primary source, verified: read the paper →

Ground Truth. (2026, July 11). Anthropic launches Claude Science, an AI workbench that keeps data in the lab and checks its own citations. Ground Truth. https://groundtruth.day/news/anthropic-claude-science-ai-workbench.html

@misc{groundtruth:anthropic-claude-science-ai-workbench,
  title  = {Anthropic launches Claude Science, an AI workbench that keeps data in the lab and checks its own citations},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/anthropic-claude-science-ai-workbench.html}
}

Topics: ai-for-science · anthropic · agents · research-tools · biology