Hands‑On Review of Claude Science: A Game‑Changer for Researchers

Earlier this year the author argued that Claude Code combined with research MCPs and Skills forms the most powerful research assistant, a paradigm he called Vibe Researching. Anthropic has now packaged this idea into an independent application named Claude Science, marketed as an AI workstation for researchers.

Research workflows are often fragmented across dozens of tools—PubMed for literature, Jupyter for analysis, R for plotting, SSH for cluster training—each with its own schema and file formats. Claude Science aims to fuse these steps into a single environment that runs wherever you work, whether on a local macOS/Linux machine or on remote HPC nodes via SSH.

The system centers on a coordinating Agent equipped with more than 60 pre‑configured Skills and connectors covering genomics, single‑cell, proteomics, structural biology, and cheminformatics. This top‑level Agent can spawn child Agents, invoke user‑written expert Agents, and reuse the same orchestration model as Claude Code Subagents, merely re‑skinned for scientific domains.

Two design choices stand out. First, every output is reproducible and traceable: when Claude generates a figure it also supplies the underlying code, execution environment, and a natural‑language explanation, all stored in the conversation history for later review or modification. Second, a dedicated reviewer Agent monitors the work, checking citation accuracy, numerical traceability, and consistency between figures and code, effectively acting as an actor‑critic pair that catches AI‑generated hallucinations.

On the compute side, Claude Science can plan large tasks such as protein folding or genome‑scale pipelines, confirm the plan with the user, and then submit jobs to existing lab resources via SSH or to Modal GPU instances. It scales from a single GPU to hundreds, keeps data on the user’s machine, and integrates models like NVIDIA BioNeMo, Evo 2, Boltz‑2, and OpenFold 3. Custom pipelines can be saved as reusable Skills, mirroring the author’s earlier Kaggle‑automation workflow with Claude Code.

Currently the app is available for macOS and Linux; any Claude subscription works, and academic groups can obtain discounted seats. However, access is limited by account bans, network restrictions, and a focus on life‑science and bioinformatics, making other domains less well‑supported.

Installation is straightforward: download the app from https://claude.com/product/claude-science, double‑click to install, launch the local service (which opens a browser on port 8765), log in, create a research project, and issue tasks—e.g., a “type‑2 diabetes target drug accessibility analysis.”

The deeper takeaway is that Claude Science’s value lies in its underlying paradigm: a central Agent, a library of domain‑specific Skills, spawnable sub‑Agents, a reviewer Agent, forkable conversations, and connectors to external tools. Using Claude Code with custom Skills, anyone can recreate a vertical AI assistant for their own field. The author warns that building separate SaaS products on top of such AI platforms faces a very short window, as major model updates or product releases from leading AI vendors can quickly render niche solutions obsolete.