Anthropic Ships an AI Workbench for Wet Lab and Computational Scientists: A Deep Look at How Claude Science Fuses Compute, 60 Plus Databases, and Reviewer Agents into One Environment
🎧 Podcast episode for paid subscribers only. Also available on Spotify.
To listen to paid episodes in Apple or Spotify, link your Substack subscription via the show settings on those platforms (instructions inside the Substack app under Subscriptions → Podcast).
Abstract
1. Anthropic launched Claude Science on June 30, 2026, a desktop AI workbench for scientists, available in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans.
2. The app collapses the fragmented research stack (PubMed, Jupyter, R, cluster terminals, database portals) into one environment with a coordinating agent, more than 60 curated skills and connectors, and specialist sub agents.
3. Every artifact ships with its full provenance: the exact code, the environment, a plain language description, and the entire message history, so figures and numbers can be reproduced or defended months later.
4. A background reviewer agent runs an actor critic loop, flagging and self correcting bad citations, untraceable numbers, and figures that do not match their underlying code.
5. It manages compute from a single laptop up to hundreds of GPUs over SSH on an HPC cluster or through Modal, keeps data on the lab’s own infrastructure, and connects to NVIDIA BioNeMo models like Evo 2, Boltz-2, and OpenFold3.
6. Early adopters include Manifold Bio and the Allen Institute, with reported wins like a year long RNA-seq contaminant found in one session and computational reviews cut from two years to weeks.
Table of Contents
The Pitch and Why It Lands
The Fragmented Stack Problem It Is Trying to Kill
Reproducibility as a First Class Citizen
The Reviewer Agent and the Actor Critic Loop
Compute Orchestration from Laptop to Cluster
The Database and Model Layer
Skills, Connectors, and the Extensibility Story
What the Early Labs Actually Did
The Business and Distribution Angle
The Skeptic Corner
Where This Fits in the Bigger Picture
The Pitch and Why It Lands
Anyone who has spent time near a computational biology lab knows the dirty secret of modern science. The glamorous part, the hypothesis and the discovery, is maybe five percent of the work. The other ninety five percent is janitorial. It is wrangling file formats nobody documented, remembering which flavor of a database uses which schema, babysitting a job on the cluster to see if it died at hour nine, and reconstructing six months later exactly how a figure got made when a reviewer asks. Anthropic looked at that mess and decided to ship a product aimed squarely at it. On June 30, 2026, the company launched Claude Science, which it describes as an AI research workbench. It is a downloadable app for macOS and Linux, and it is in beta for Pro, Max, Team, and Enterprise plans.
The reason the pitch lands is that it is not another chat window bolted onto a search box. The framing is that a scientist works inside a single running environment where all the stages of research happen. You analyze the literature, you run multi step analyses, you generate figures and manuscripts, and you iterate on them in plain language until they are ready to submit. The differentiator, and the thing that will matter most to the technical crowd, is that every output carries an auditable history of how it was made. That is the whole game. In science, an answer you cannot reproduce is not an answer, it is a rumor. Anthropic clearly understands that the bar for this audience is not fluency, it is defensibility.
