Claude Science: An AI Workbench for Research

June 30, 2026, a Tuesday: Anthropic shipped Claude Science in beta. Sixty-plus pre-configured skills, direct access to more than 60 scientific databases, one generalist AI agent, and a separate reviewer agent checking its work. As of August 2026 it runs on macOS and Linux, is usable remotely over SSH on HPC clusters, and comes bundled with an existing Claude Pro, Max, Team, or Enterprise plan — there's no separate pricing tier, according to the product page.

The numbers sound like a lot, but the underlying idea is simple: the browser tabs, downloads, notebooks, and terminal sessions a researcher juggles all day get pulled into one workspace, coordinated by a single agent.

The problem: scattered tools, one researcher

A day in computational biology tends to look like this: one tab open to UniProt to look up a protein, another pulling a structure file from PDB, a third scanning ClinVar for variant records, then a terminal window to queue an analysis job on an HPC cluster. Every tool has its own interface, its own API, its own learning curve. A genomicist and a cheminformatician barely touch the same databases, but they share the same friction — more time spent switching tools than producing results.

Claude Science collapses that friction into one workspace. The concept itself isn't new — an orchestration layer that pulls databases, code, and compute into a single interface, coordinated by a generalist AI agent. Anthropic is explicit that this is a workbench, not a new foundation model: an orchestration layer over existing tools, data, and compute. That distinction matters — Claude Science isn't doing new science, it's making existing scientific infrastructure easier to reach.

A researcher's query into that workspace might look like this:

Pull the UniProt entry for human TP53, match it against known structures in
PDB, list pathogenic variants from ClinVar, and render the result as a
genome browser track.

That single request would otherwise mean four separate tools, four different input methods, and probably an afternoon.

60+ skills: expert tools for non-experts

At the core of Claude Science are 60-plus curated "skills" pre-configured for genomics, single-cell biology, proteomics, structural biology, and cheminformatics. Those skills query more than 60 scientific databases directly — UniProt, PDB, Ensembl, Reactome, ClinVar, and ChEMBL among them. On the structural-biology side, Anthropic's announcement notes that Claude Science integrates with NVIDIA's BioNeMo Agent Toolkit, giving it access to models like Evo 2, Boltz-2, and OpenFold3.

"Skill" here means an instruction package that encapsulates a database or tool's API, query syntax, and output format. In practice, that means a researcher doesn't need to memorize ChEMBL's REST API or Ensembl's query language — the skill already carries that knowledge. That's also where the power and the risk sit side by side: getting expert-level access to a tool you're not an expert in is genuinely useful, but it can also mean you lack the expertise to evaluate what comes out of it.

This table isn't exhaustive — 60-plus databases and skills go well beyond it — but it shows how the five domains map to different specialties. What a cell biologist needs from GEO barely overlaps with what a medicinal chemist needs from ChEMBL; Claude Science's pitch is giving both the same interface.

A reviewer agent and an auditable history

A separate reviewer agent runs on top of the generalist agent. Its job is to check citations and calculations, flagging and correcting errors — an accuracy layer added on top of whatever the generalist agent produces. That two-agent split matters: a single agent that both generates and grades its own work risks missing its own mistakes; a reviewer operating in a separate context reduces that risk.

Claude Science also keeps auditable histories of its outputs. A researcher can trace back how a result was produced — which database was queried, what computation ran, which intermediate steps were taken. In scientific work, that's a concrete answer to the "black box" critique: the output isn't just text, it comes with a traceable chain of how it got there.

The product also renders scientific artifacts directly — 3D protein structures, genome browser tracks, chemistry drawings — so results can be evaluated inside the workspace instead of being ported to a separate visualization tool.

Realistic limits: validation still matters

Claude Science is still beta software, and that has practical consequences. Anything headed into a paper or a regulatory submission needs human validation — that's not a suggestion, it's a real requirement. The reviewer agent tries to catch errors, but automated checking doesn't replace domain expertise.

Here's the honest tension: the same 60-plus skill package that makes the product powerful is also its fragility. These skills package domain tools for people who aren't experts in every one of those tools — that's genuinely useful, because a computational biology grad student no longer needs to memorize ChEMBL's API by heart. But it also means outputs need domain-expert review rather than blind trust. If correctly interpreting a tool's output requires already knowing that tool, the ease of access a skill provides can produce a false sense of confidence right alongside the convenience.

My honest take: this isn't a flaw specific to Claude Science, it's a reflection of what doing science actually requires. No tool — AI-assisted or not — replaces scientific rigor; what Claude Science's auditable-history and separate-reviewer design does is build the right scaffolding to make that rigor easier to practice.

Who actually benefits

Wet-lab and computational biology labs are the clearest winners, especially teams that move constantly between databases. For biotech and pharma startups the value proposition is a bit different: small teams usually can't hire a dedicated specialist for every database, and Claude Science fills that gap to some degree. For grad students, the benefit shows up mostly as time saved — a week that would have gone to learning one database's API turns into actual analysis instead.

Is this for you?

If your work regularly means switching between multiple scientific databases, stitching together data in different formats, and repeating query-download-analyze loops, Claude Science is worth trying. If you work deeply within a single database that requires real specialization — and you already know its interface well — the payoff is more limited. Either way, the thing to remember is the same: this is a workbench, not a referee. The final word still belongs to a domain expert.

If you're curious how Claude's skill system works more broadly outside of research, our piece on what Claude Skills are is a good primer, and our 2026 Claude model guide can help you pick the right model for your workload. If you're interested in how agents get reliable context to work with, our context engineering for AI agents piece helps make sense of Claude Science's reviewer-agent design too. For the security side of skill packages, see our Claude skill and plugin security scanning writeup. More AI coverage lives on our AI category hub.

Frequently Asked Questions

How do I get access to Claude Science?

There's no separate charge — Claude Science is bundled with an existing Claude Pro, Max, Team, or Enterprise plan. It currently runs on macOS and Linux, and is also usable remotely over SSH on HPC clusters.

How much can I trust Claude Science's output?

A separate reviewer agent checks citations and calculations, and an auditable history is kept for every output, but that doesn't replace human validation. Anything going into a paper or a regulatory submission still needs review by a domain expert.

Is Claude Science a new AI model?

No. Anthropic explicitly describes it as a workbench — an orchestration layer over existing databases, tools, and compute — not a new foundation model.

Which scientific domains does it cover?

As of now, it offers 60-plus skills focused on genomics, single-cell biology, proteomics, structural biology, and cheminformatics, and can directly query more than 60 databases, including UniProt, PDB, Ensembl, Reactome, ClinVar, and ChEMBL.