# Claude Science Claude Science is [[Anthropic]]'s "AI workbench for scientists": a separate desktop app where [[Claude]] runs analyses, searches scientific databases, manages compute, and drafts figures and manuscripts. It launched in public beta on 2026-06-30 for Pro, Max, Team and Enterprise plans. Why it exists: research work is scattered. A single question can mean PubMed, a dozen databases with their own schemas, Jupyter, R, a cluster terminal and some bespoke file viewer. Anthropic's bet is to put all of that behind one conversation, and to make every result traceable back to the code that produced it. There's no new model behind it (the FAQ says so explicitly). It uses the same Claude models your plan includes; the launch video shows [[Claude Opus 4.8]] in the model picker. What's new is everything around the model: tools, database connectors, compute integrations and the provenance layer. ## How it works - **Runs where your data lives.** You install it on your laptop, a lab Linux box, an HPC login node or a cloud VM, then connect from your browser. Raw datasets stay on your machines; only the context needed for each step goes to Claude. Code (Python, R, shell) runs in a sandbox, and you approve each new folder, network host and remote job. - **A team of agents.** A generalist coordinating agent can spin up specialist agents (and ones you define). In the launch video, a literature review request fans out into five parallel retrieval agents (bioRxiv, OpenAlex, PubMed, embedding methods, CELLxGENE). - **Provenance on every artifact.** Each figure, table or notebook ships with the exact code, the environment, a plain-language description of how it was made and the full conversation. You can fork a session to compare two approaches without losing the original thread. - **A background reviewer.** It flags incorrect citations, numbers it can't trace back to evidence, and figures that don't match their code. The docs are careful here: it checks claims against the execution record and doesn't re-run analyses, so it reduces errors without eliminating them. Claude Science is also not meant for clinical or diagnostic use. - **Native scientific renderers.** 3D protein structures, genome browser tracks, alignments, chemical structures, PDFs, plus Markdown and LaTeX previews for manuscripts. You can annotate a figure and ask for a change in plain words (e.g. log scale on an axis); the agent edits its own code. - **Compute management.** Persistent Python and R kernels keep data in memory across the analysis. For heavy jobs it writes batch scripts and submits them over SSH to your own workstation or Slurm cluster, or to cloud GPUs through your own Modal account (billed by Modal, not Anthropic). The video shows the approval prompt before a Modal job runs on H100 GPUs. - **Domain-ready, extensible.** Pre-configured for genomics, single-cell, proteomics, structural biology and cheminformatics, with 60+ curated skills and connectors (UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO...) and NVIDIA's BioNeMo models (Evo 2, Boltz-2, OpenFold3). You can save any pipeline as a reusable [[AI Agent Skills|skill]] or plug in your own tools through connectors ([[Model Context Protocol (MCP)]]); future sessions inherit them. The docs say it isn't limited to the life sciences, but the featured connectors clearly are. ## Availability - Public beta. Included in Pro, Max, Team and Enterprise (not Free). On Team and Enterprise, an admin has to enable it first. - Usage counts toward the same limits as the rest of your plan, [[Claude Code]] and [[Claude Cowork]] included. - The launch post said macOS and Linux; the product page and docs now list macOS 13+, Windows 11 (x64) and Linux x64 (glibc; needs bubblewrap and socat). About 5 GB of disk. - A discounted Team plan exists for scientists at academic and nonprofit institutions. - At launch, Anthropic also opened an AI for Science program: up to 50 projects, up to $30,000 in credits each, plus up to $2,000 of Modal compute for some (applications closed 2026-07-15; projects run 2026-09-01 to 2026-12-01). ## Early results Anthropic reports - Jérôme Lecoq (Allen Institute) built a review pipeline of about 20 custom skills with writer and reviewer agent pairs. A long-form review used to take his team up to two years; he now has about 10, many over 100 pages. - Stephen Francis (UCSF) ran germline analyses on glioma in roughly a tenth of the previous time, and his group validated the results independently. - Manifold Bio used it to nominate targets for its latest experiments, ranking candidates against criteria learned from its own proprietary data. ## Reception The Hacker News launch thread (564 points, 174 comments) was split. The skeptics worried about a wave of paper-mill publications, pointed out that this is mostly the *data science* part of science, and that the connectors are almost all biology and pharma. One tester reported wrong DOIs and hallucinated references during a literature review. On the positive side, a former All of Us engineer noted that the local server plus browser UI fits locked-down Trusted Research Environments, where desktop apps can't run but JupyterLab-style tunnels can. Someone who built one of the launch connectors argued that forcing good LLM-facing APIs onto scientific databases (many are still FTP-only!) is valuable on its own. ## My take For me, the most interesting part is the provenance plus reviewer combination. Agents produce a lot of output fast, and the bottleneck becomes "can I trust and defend this?". Welding code, environment and conversation to every artifact, then having a second agent check claims against what actually ran, is the right answer to that problem. And it applies far beyond science: any knowledge work done with agents needs the same audit trail. The HN skeptics have a point, though. A reviewer that checks claims against the execution record still can't tell you whether the question was a good one. Rigor stays the scientist's job. ## References - Launch video (Claude, 2026-06-30, 1:26, music only; content read from the on-screen demo and the description): https://www.youtube.com/watch?v=idtMsa_1yNk - Announcement: https://www.anthropic.com/news/claude-science-ai-workbench - Product page and FAQ: https://claude.com/product/claude-science - Documentation: https://claude.com/docs/claude-science/overview - Team plan for scientists: https://claude.com/programs/team-plan-for-scientists - Hacker News discussion: https://news.ycombinator.com/item?id=48735770 ## Related - [[Claude]] - [[Anthropic]] - [[Claude Code]] - [[Claude Cowork]] - [[AI Agents]] - [[AI Agent Skills]] - [[Model Context Protocol (MCP)]] - [[Data Science]] - [[AlphaFold]]