{"slug": "claude-science-anthropics-ai-workbench-with-60-research-tools", "title": "Claude Science: Anthropic’s AI Workbench with 60 Research Tools", "summary": "Anthropic launched Claude Science on June 30, 2026, a desktop workbench providing a Claude agent with access to 60+ scientific databases, sandboxed compute, and HPC clusters, with a provenance system logging every analysis step. The tool is available to all paid Claude subscribers and is already in production at UCSF, Allen Institute, and Manifold Bio, enabling reproducible research workflows.", "body_md": "Anthropic launched [Claude Science](https://www.anthropic.com/news/claude-science-ai-workbench) on June 30, 2026 — a desktop workbench that gives a Claude agent access to 60+ scientific databases, sandboxed compute environments, and HPC clusters, wrapped in a provenance system that logs every step of every analysis. It is not a new model. While OpenAI fine-tunes GPT-Rosalind and Google DeepMind ships proprietary science models, Anthropic bet on workflow. That bet is now live for all paid Claude subscribers, and it is already in production at UCSF, Allen Institute, and Manifold Bio.\n\n## A Workflow Layer, Not a New Model\n\nThe architecture is a coordinating agent that spawns specialist sub-agents via [60+ curated Skills](https://techcrunch.com/2026/06/30/anthropics-claude-science-bets-on-workflow-not-a-new-model-to-win-over-scientists/), built on the same Skills and Model Context Protocol primitives that power Claude Code. The pattern runs four steps: Discovery → Invocation → Artifact Generation → Verification. After every run, a reviewer agent independently checks citations and calculations against execution records. You do not configure this pipeline — you describe what you want in natural language, and Claude Science routes it.\n\nThe competitive contrast is worth stating plainly. OpenAI’s approach is a fine-tuned model in gated enterprise preview. Google DeepMind’s approach is proprietary foundational models on the Gemini for Science platform. Anthropic’s approach is a workflow layer on top of existing Claude models — available immediately to Pro, Max, Team, and Enterprise subscribers, no new pricing tier required for most users.\n\n## Reproducibility Was Always the Blocker\n\nThe reason most research labs kept LLMs at arm’s length was not capability — it was auditability. If you cannot explain exactly what the model did, you cannot publish it, satisfy an IRB, or defend it in a regulatory review. Claude Science addresses this directly: every artifact includes the exact source code that produced it, the runtime environment and package versions, a plain-language methodology description, and the full message history. That bundle travels with every output.\n\nEarly adoption reflects this. UCSF Brain Tumor Center accelerated glioma germline genetic workups roughly 10x with results independently validated by the lab team. Allen Institute neuroscientist Jérôme Lecoq built approximately 20 custom skills for a computational review pipeline that compresses two-year literature reviews into weeks. Manifold Bio used Claude Science end-to-end for drug target nomination in tissue-specific medicines, assessing surface expression, trafficking, and safety criteria per candidate. These are not demos — they are production workflows in regulated research environments.\n\n## 60+ Skills, Custom Extensions, BioNeMo\n\nThe skill library covers the major scientific databases: UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, and the NVIDIA BioNeMo toolkit (Evo 2 for genomics, Boltz-2 for protein structure, OpenFold3). The execution environment is sandboxed Python with NumPy, pandas, SciPy, and matplotlib — plus R with tidyverse and ggplot2. Performance benchmarks from Anthropic include 1.3M-cell preprocessing dropping from 52 minutes to 25 seconds, and cheminformatics searches running up to 3000x faster.\n\nFor developers, the custom skills interface is the important part. Because [Claude Science](https://claude.com/product/claude-science) is built on the same MCP primitives as Claude Code, the extension model is already familiar. You can build domain-specific tools on the same infrastructure and deploy them as specialist sub-agents within a session. The macOS and Linux-only constraint at launch — Windows is not supported — creates a clear gap that developers can fill with web frontends or API wrappers.\n\n## Compute Without the Configuration\n\nClaude Science handles resource scaling from a laptop to hundreds of GPUs. SSH integration enables SLURM job submission to HPC clusters. Modal provides on-demand GPU compute. The agent reasons about what compute a given task requires rather than needing pre-configured routing — a meaningful distinction from tools where you manually select environments. Sensitive data stays on local systems; only essential context is sent to Claude. The explicit approval workflow (user approves, app requests permission, code runs in an OS-level sandbox, reviewer checks) gives regulated environments the audit trail they need.\n\n## Two Paths Forward, One Deadline\n\nAn open-source alternative arrived five days after Claude Science. [OpenScience](https://github.com/ai4s-research/open-science) from Synthetic Sciences (YC W26) is Apache 2.0, model-agnostic — Claude, GPT, Gemini, local models — and ships with 250+ editable skills and 30+ database connectors. It is less polished and more open, a reasonable trade-off if vendor lock-in is a concern. The comparison between the two platforms is worth reading if you are evaluating which fits your lab or stack.\n\nOne concrete near-term action: Anthropic’s Claude Science Grants program offers $30,000 in compute credits, and applications close July 15. If you are building research tooling, that deadline is in two days.\n\n## Key Takeaways\n\n- Claude Science is a workflow layer, not a new model — it runs existing Claude models with 60+ scientific skills on top\n- The reproducibility system (source, environment, methodology, message history bundled with every artifact) is the feature that unlocks adoption in regulated research\n- Real-world results: 52-minute preprocessing to 25 seconds, ~10x faster germline workups at UCSF, 2-year literature reviews compressed to weeks at Allen Institute\n- Developers can build custom skills using the same MCP primitives as Claude Code — no new toolchain required\n- OpenScience is a model-agnostic open-source alternative (Apache 2.0) for teams that want to self-host\n- Claude Science Grants ($30K in credits) close July 15", "url": "https://wpnews.pro/news/claude-science-anthropics-ai-workbench-with-60-research-tools", "canonical_source": "https://byteiota.com/claude-science-ai-workbench/", "published_at": "2026-07-13 03:13:54+00:00", "updated_at": "2026-07-13 03:15:29.750647+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-tools", "ai-infrastructure", "ai-research"], "entities": ["Anthropic", "Claude Science", "UCSF", "Allen Institute", "Manifold Bio", "OpenAI", "Google DeepMind", "NVIDIA BioNeMo"], "alternates": {"html": "https://wpnews.pro/news/claude-science-anthropics-ai-workbench-with-60-research-tools", "markdown": "https://wpnews.pro/news/claude-science-anthropics-ai-workbench-with-60-research-tools.md", "text": "https://wpnews.pro/news/claude-science-anthropics-ai-workbench-with-60-research-tools.txt", "jsonld": "https://wpnews.pro/news/claude-science-anthropics-ai-workbench-with-60-research-tools.jsonld"}}