How Claude Science Works (Summary)

Source: Anthropic product pages, launch blog, documentation and partner blogs (Modal), independent coverage (PaperBanana). This report summarizes architecture, features, integrations, and practical workflow for Claude Science — Anthropic's AI workbench for scientific research (beta, June 30, 2026).

Quick overview

• Claude Science is an application (workbench), not a new model. It uses existing Claude models (Sonnet family, Fable, Mythos depending on plan).
• Purpose: bring literature search, data analysis, compute orchestration, and manuscript/figure generation into a single reproducible environment for scientists.
• Runs locally (macOS, Linux) or on remote infrastructure (SSH/HPC, Modal). Data stays on the user's infrastructure; only necessary context is sent to Claude.

Core components

Agents and Skills

- Coordinating agent: generalist that manages the research session, spawns sub-agents for specialist tasks.

- Skills: folders of instructions, scripts, environment definitions and connectors that encapsulate pipelines (e.g., single-cell QC). Users and Anthropic publish domain-specific skills.

Connectors & Model Context Protocol (MCP)

- Connectors provide direct access to tools and databases (Benchling, PubMed, BioRender, 10x Genomics, Synapse, Wiley Scholar Gateway, ClinicalTrials.gov, UniProt, PDB, etc.).

- MCP allows Claude to reference original sources; outputs include numbered citations that trace back to the source records for provenance and auditability.

Compute orchestration

- Claude Science manages environments and compute: local kernels, SSH to HPC, or cloud compute via Modal.

- Scales from a single GPU up to hundreds. Modal integration provides elastic containers, shared volumes, and reproducible images for each compute step.

Reviewer / actor-critic agents

- A reviewer agent inspects outputs, checks citations and calculations, flags mismatches (e.g., figure not matching code) and can self-correct or surface issues to the user.

Reproducibility & artifacts

• Every artifact (figure, table, notebook) includes the exact code, environment, and conversation that produced it, so results are reproducible and editable.
• Built-in renderers for 3D protein structures, genome tracks, chemical structures, and more.
• Plain-language figure iteration: request edits in natural language and Claude rewrites and reruns code to produce updated figures.

Typical workflows

1. Literature review: specialist agents query PubMed, bioRxiv, preprints; reviewer agent checks citations; outputs saved to an evidence-state DB.
2. Data analysis: upload dataset or connect to institutional storage; Claude builds environment, runs analysis (Python/R kernels), and produces figures with provenance.
3. Compute-heavy tasks: offload to Modal or lab HPC via SSH; Claude manages job submission and resource scaling.
4. Manuscript drafting: draft inside app; figures and code bundled; reviewer agent verifies citations and numbers.

Limitations & caveats

• Agent-checked citations reduce risk of hallucinations, but they are not a replacement for human verification.
• Product is new (beta) and feature details may change; many claims are self-reported by Anthropic and partners.
• Currently macOS / Linux only; plan-gated (Pro/Max/Team/Enterprise).

References & further reading

- Anthropic: Claude for Life Sciences (Oct 20, 2025)
- Anthropic: Claude Science launch blog (Jun 30, 2026)
- Claude product page: /product/claude-science
- Modal: Modal integration brings scalable compute to Claude Science
- PaperBanana: "Claude Science: What Researchers Need to Know in 2026"
  

Generated by an automated web research script. For more detail or a slide/diagram pack, ask me to expand this report (add diagrams, sequence diagrams, or step-by-step demo instructions).