Claude Science Explained: Inside Anthropic's 60-Skill AI Research Workbench

On June 30, 2026, Anthropic quietly shipped one of its most architecturally interesting products yet: Claude Science, a beta AI workbench built for scientists but powered by the same MCP-and-skills stack that's reshaping how every developer builds with Claude. If you've been following Claude Code's shift toward Skills and Subagents, Claude Science is the clearest real-world proof that the pattern scales — Anthropic used it to wire Claude into more than 60 specialized research tools without writing 60 custom integrations.

That's the part worth paying attention to, even if you never touch a genome browser. Claude Science is a case study in composable agent design, and it's a preview of where Claude's "skills economy" is headed next.

What Claude Science Actually Is

Claude Science is a standalone desktop app — available in beta on macOS and Linux — that consolidates the fragmented tooling scientists normally juggle across a dozen apps, notebooks, and command-line packages into one Claude-driven workspace. It's rolling out to Pro, Max, Team, and Enterprise plans, though Team and Enterprise admins need to explicitly enable it for their org.

Instead of a general-purpose chat interface, Claude Science ships pre-loaded with domain packages for:

• Genomics and single-cell analysis
• Proteomics
• Structural biology
• Cheminformatics

Under the hood, it connects to the NVIDIA BioNeMo toolkit, including models like Evo 2, Boltz-2, and OpenFold3, and can hand off compute-heavy jobs to laptops, HPC clusters, or on-demand GPUs depending on what a task needs.

None of that is exotic capability by itself — researchers have had access to these tools for years. What's new is that Claude now orchestrates them as a single conversational layer, deciding which tool to call, running it, and explaining the result in plain language.

The Architecture: 60+ Skills, One Agent

This is the piece that matters if you're a developer, not a biologist. Claude Science's "over 60 curated skills and connectors" aren't 60 bespoke API integrations Anthropic hand-built for this one app. They're built on the same Skills and MCP (Model Context Protocol) primitives available in Claude Code today — the same system we walked through in our custom Skills tutorial.

A Skill, in Anthropic's framework, is a self-contained folder of instructions, scripts, and resources that Claude loads on demand rather than keeping in context permanently. MCP servers expose external tools and data sources through a standard protocol Claude already knows how to call. Claude Science's real innovation is treating an entire scientific field as a library of pre-packaged skills rather than a monolithic app:

Discovery — Claude inspects the task and decides which skill (say, a protein-folding connector) is relevant.

Invocation — The relevant MCP connector runs, often on external compute, without loading unrelated tooling into context.

Artifact generation — Results come back as auditable artifacts: 3D protein structures, genome tracks, or chemical structures rendered natively, each traced back to the exact code and environment that produced it.

Verification — A built-in reviewer agent checks citations and calculations before the result is presented, and users can fork a session to compare two analytical approaches side by side.

If that discovery → invocation → verification loop sounds familiar, it should — it's the same pattern we described in our guide to Claude Code Subagents, just applied at the scale of an entire scientific discipline instead of a single coding task.

Why Reproducibility Is the Real Headline

Most AI research tools optimize for "give me an answer fast." Claude Science was clearly built around a different constraint: every output has to be independently verifiable. When it generates a figure, it attaches:

• The exact code that produced it
• The runtime environment and package versions used
• A plain-language description of the methodology
• The full message history that led to the result

This matters because reproducibility is the single biggest blocker to using LLMs in regulated or peer-reviewed environments. A chatbot answer nobody can audit is a liability in a lab notebook. By making every artifact traceable to its source code and environment, Anthropic is directly addressing the "AI hallucination in science" problem that's kept many labs from adopting LLM tooling at all.

For developers building AI products of any kind — not just scientific ones — this is a pattern worth stealing: pair every generated output with the exact context that created it. It's the difference between a demo and a tool people can actually rely on for decisions that matter.

The AI for Science Program: How to Get Involved

Alongside the product launch, Anthropic opened applications for its AI for Science program, aimed at funding up to 50 projects in biology and biomedical research:

If you're a researcher (or know one) looking to test Claude Science on a real dataset, the credit stack is generous enough to be worth the application effort — $30K in API/product credits plus compute is a meaningful research grant, not a marketing gesture.

What This Means If You're Building With Claude (Not Doing Science)

Here's the part that's directly relevant to the developers and certification candidates in our audience. Claude Science demonstrates three things about where Anthropic is pushing the platform in mid-2026:

1. Skills are becoming the default extension mechanism. Rather than fine-tuning models per vertical, Anthropic is packaging domain expertise as loadable Skills and MCP connectors. If you're studying for the Claude Certified Architect exam, expect Skills architecture — not prompt engineering alone — to be a growing share of what "building well with Claude" means. 2. Verification agents are becoming standard practice. The built-in reviewer agent that checks citations and math before showing results to the user is a pattern you can replicate in your own multi-agent systems today, regardless of domain. Any pipeline that generates factual claims should have a dedicated verification step, not just a single generation pass. 3. Compute orchestration is now part of the agent's job. Claude Science routing tasks across laptops, HPC clusters, and on-demand GPUs shows Anthropic treating "where does this run" as something the agent reasons about, not something a human pre-configures. That's a preview of how agentic coding tools will likely handle resource-intensive tasks going forward.

Common Mistakes to Avoid With Beta Tools Like This

Since Claude Science is explicitly in beta, a few practical notes if you or your team plan to try it:

• Don't treat beta reproducibility claims as a compliance guarantee. The auditable-artifact system is a strong step forward, but beta software changes fast — verify outputs against your institution's existing standards before citing them in publications.
• Check admin enablement first if you're on a Team or Enterprise plan. Unlike Pro/Max, this isn't a self-serve toggle — you'll need an admin to turn it on, which we covered in more detail in our Claude Enterprise admin analytics guide.
• Don't assume all 60+ skills are equally mature. Curated skill libraries roll out unevenly; test the specific domain package you need (e.g., structural biology) before building a workflow around it.

Claude Science vs. Claude Code Skills: What's Different

It's easy to conflate Claude Science with "Claude Code but for scientists," but the products solve different problems. Claude Code's Skills system is a general-purpose extension mechanism — any developer can write a Skill for any domain, from database migrations to ad copywriting. Claude Science is a curated, vertical-specific deployment of that same mechanism, pre-loaded and pre-tested for one field.

That distinction matters for how you think about adopting either one:

• If you need Claude to do something novel and specific to your own workflow, you write a custom Skill using the same tutorial we linked above — that's the Claude Code path.
• If you need Claude to operate inside an established, well-instrumented domain where Anthropic has already done the integration work, a vertical product like Claude Science is the faster route — you inherit a maintained library instead of building one.

Expect more vertical spinoffs of this pattern. The same playbook — take Skills and MCP, curate a domain-specific bundle, add a verification layer, ship as a standalone app — is a template Anthropic can reuse for law, finance, or engineering just as easily as biology. If you're building a Claude-based product for a specific industry, Claude Science is worth studying as a blueprint, not just a research tool.

Frequently Asked Questions

Is Claude Science available on Windows?

Not yet. The beta currently supports macOS and Linux only; Anthropic hasn't announced a Windows release date.

Do I need a special API plan to use Claude Science?

No new billing tier — it's included with existing Pro, Max, Team, and Enterprise subscriptions. Team and Enterprise users need admin approval to enable it.

Can I build my own domain-specific version of what Claude Science does?

Yes. Claude Science is built on publicly available Skills and MCP primitives. Any developer can package a set of connectors, verification logic, and artifact rendering for their own vertical using the same Claude Code tooling — see our Skills tutorial for the starting point.

How is this different from just asking Claude to run Python code?

Claude Science adds structured skill discovery across 60+ pre-integrated tools, native rendering of scientific artifacts (3D structures, genome tracks), a dedicated verification/reviewer agent, and orchestration across local and remote compute — well beyond what an ad hoc code-execution prompt provides.

Key Takeaways

• Claude Science is a beta AI workbench (macOS/Linux, Pro/Max/Team/Enterprise) that unifies genomics, proteomics, structural biology, and cheminformatics tooling behind a single Claude interface.
• It runs on the same Skills and MCP architecture available in Claude Code — over 60 curated connectors, not custom one-off integrations.
• Reproducibility is the core design principle: every artifact ships with its source code, environment, and methodology attached.
• Anthropic's AI for Science program is funding up to 50 research projects with up to $30K in credits through a July 15, 2026 application deadline.
• The underlying patterns — skill discovery, verification agents, compute orchestration — are directly applicable to any Claude-based product you're building today.

Next Steps

Want to build the same kind of Skills-based architecture Claude Science demonstrates? Start with our hands-on Claude Code Skills tutorial to build your first custom Skill, then study the multi-agent verification pattern in our Subagents guide.

If you're preparing for the Claude Certified Architect (CCA) exam and want practice questions that cover exactly this kind of Skills and MCP architecture, check out AI for Anything's CCA practice test bank — built to keep pace with releases like this one as they land.

---

Sources: Anthropic — Claude Science announcement, MarkTechPost — Anthropic Launches Claude Science Beta