Last year Anthropic shipped Claude Code, and software engineers got a development environment where the model was so tightly integrated into the edit-run-debug loop that context-switching between the AI and the terminal mostly stopped being a thing. The pattern was deliberate: find the most painful friction in an expert workflow and collapse it.

Claude Science, which launched in beta on July 1, is the same pattern applied to scientific research.

Desktop app. Runs Python, R, and shell. Connects to your HPC cluster over SSH, or Modal for on-demand compute at scale. Sixty-plus curated skills and connectors pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Every output carries an auditable history of how it was produced, so a collaborator or peer reviewer can trace the path from raw data to published figure.

Multiple outlets reached for the same phrase: "Claude Code for scientists." That's the positioning and it's accurate.

What the reviewer agent actually does

Here's the part that's different from just Claude running Jupyter notebooks.

Claude Science includes a reviewer agent that runs after analysis is complete. It checks citations against the literature, flags numbers that don't appear in the cited sources, and identifies figures whose visual output doesn't match the code that generated them. When it finds a mismatch, it self-corrects.

That's a meaningful design choice, not a convenience feature. The reproducibility crisis in academic research is mostly a citations-and-methodology problem. Bad papers fail because figures get visually edited after generation, citations get misread six months after the researcher read them, and statistical procedures aren't clearly connected to their inputs. A reviewer agent that runs those checks automatically is attacking the actual failure mode.

The grant program: Anthropic is funding up to 50 Claude Science AI for Science projects with up to $30,000 in Claude API credits per project. Modal is adding up to $2,000 in compute credits for selected projects. Applications close July 15, with award notifications by July 31. Projects run September 1 through December 1, 2026. Focus is biomedical and biology research, specifically at the postdoctoral and graduate level.

Source spread

• Anthropic — Claude Science, an AI workbench for scientists — [hype]. Primary source; useful for feature list, grant details, and the 60+ skills claim. No independent performance data.
• WorthvieW — Claude Code of lab research — [builder]. Good independent framing of the Claude Code parallel and workflow implications.
• Mervin Praison — Claude Science Explained — [builder]. Best breakdown of the auditable artifact system and its reproducibility implications.
• ChatForest — Claude Code of lab research guide — [skeptic]. Notes the 60+ skills skew genomics-heavy; physics, chemistry, and social science coverage is thinner in the first release.

Pros & cons

What's real:

• The reviewer agent is the right design choice. Building validation into the workflow rather than leaving it to the researcher is a structural improvement. The specific failure modes it catches — citation mismatches, figures inconsistent with their underlying code — are exactly the failure modes that drive retractions.
• The auditable artifact history is the correct answer to the trust problem in AI-assisted science. Peer reviewers have no mechanism to accept AI-assisted papers without provenance. Claude Science's built-in history gives them one.
• The 60+ curated scientific skills are not just a marketing number. Pre-configured genomics and single-cell connectors mean researchers skip the setup cost of wiring up domain-specific tools. That's the actual time sink for most computational researchers starting a new analysis.
• Modal integration means the compute ceiling is cluster-scale, not laptop-scale. Running a large protein-structure or single-cell atlas analysis doesn't require the researcher to have an HPC allocation already.

What deserves a side-eye:

• Beta, and Anthropic's betas have tended to mean "works in demos, narrower in practice." The skill library is genomics-heavy; a physicist or materials scientist will find fewer pre-built connectors at launch.
• The reviewer agent checks citations against accessible literature. Pre-print servers, internal lab datasets, and proprietary databases are outside that scope by default. Researchers working with novel unpublished data have a gap the reviewer agent won't cover.
• The grant is $30k in Claude API credits, not unrestricted research funding. Projects are implicitly selected for compute-heavy, Claude-Science-native workflows. That's a real filter on what research gets funded here.

What builders need to know

Further reading

• Anthropic — Claude Science, an AI workbench for scientists — official launch post; feature list, grant overview, domain coverage, Modal partnership
• Anthropic / Claude Help Center — AI for Science grant program — grant eligibility, application process, deadline, award timeline
• WorthvieW — Claude Code of lab research — independent analysis of the Claude Code parallel and what it means for workflow
• Mervin Praison — Claude Science explained — best walkthrough of the auditable artifact system and reproducibility implications
• ChatForest — builder guide — independent read including the skill-coverage gaps in the first release