This week, Anthropic was released cloud science. This is an app for scientists, available in beta. It runs on Anthropic’s existing cloud model, not a new model. The app targets researchers who manage databases, notebooks, and cluster terminals. It runs multi-step research and records how each result was generated. The beta is available for Pro, Max, Team, and Enterprise plans.
Cloud Science builds on Anthropic’s life sciences work from last fall. That earlier work connected the cloud to the scientific ecosystem through MCPs and skills.
What is cloud science?
Cloud Science is an AI workspace for research. It integrates the tools and packages that researchers most commonly use. It analyzes literature, performs multi-step research, and prepares detailed artifacts. You can refine data and manuscripts until they are ready for publication.
You speak to a generalist coordination agent in clear language. That agent has access to over 60 curated skills and connectors. These are pre-configured for genomics, single-cell, proteomics, structural biology and cheminformatics.
You can run it locally on macOS or Linux. You can also work on a remote machine over SSH or on an HPC login node. Each output contains an audible history of how it was created.
How does multi-agent architecture work
A generalist coordination agent receives your request in simple language. It may appoint other agents to handle the work. It can also engage expert agents that users create themselves. NVIDIA describes these as pre-configured, domain-specialized agents. Everyone knows the established workflow for their area.
A separate reviewer agent runs as the pipeline executes. It inspects the output step by step. It flags misquotes and numbers that it can’t figure out. It also flags data that does not match their underlying code. Then as it progresses, it corrects itself.
Reproducibility and provenance
Scientific research is inherently visual. So Cloud Science produces figures and manuscripts along with the code that created them. It natively presents 3D protein structures, genome browser tracks, chemical structures, and more.
When it generates a statistic, it records the exact code and environment. It also records details and complete message history in a simple language. This makes it easier to verify work and reproduce months later.
You can edit figures in simple language. For example, you can tell it to change the axes to log scale. The agent then edits its code. You can also organize a session to compare the two approaches without losing the original.
calculate that scale on demand
Larger analyzes often require more than a laptop. Folding of proteins is an example of this. Cloud science involves drafting a plan before accessing new resources. It asks for approval and lets you review or cancel any decisions. It then writes the job and submits it to your own infrastructure.
This means your HPC cluster is over ssh or your model account. Analysis can be scaled from a single GPU to hundreds as needed. Because agents keep the context in memory, a large dataset is loaded only once.
The app runs on your lab’s own infrastructure. So large or sensitive datasets never have to leave your existing system. Only the necessary context for each step is sent to the cloud.
Domain Coverage and NVIDIA BioNeMo
Scientific knowledge is scattered across hundreds of specialized sources. In biology, this includes UniProt, PDB, Ensembl, and Reactome. It also includes ClinVar, ChEMBL, GEO, journal and preprint servers. Expert agents interrogate and synthesize these sources for you.
Cloud Science also utilizes the skills of NVIDIA’s Bionemo Agent Toolkit. The toolkit packages GPU-accelerated capabilities as callable skills. It connects seamlessly to Evo 2, Boltz-2, and OpenFold3. The Evo 2 is a genomics foundation model. Boltz-2 handles biomolecular interaction prediction. OpenFold3 handles protein structure prediction.
use cases with examples
Beta users have run single-cell RNA sequencing analysis and CRISPR screen design. He has also conducted protein structure prediction and cheminformatics.
- target enrollment: Manifold Bio designs tissue-targeted drugs. It used cloud science to nominate targets for its latest experiments. For each tissue and target, the app assessed surface expression, trafficking, and safety. The candidates are then ranked against Manifold’s own proprietary criteria. Manifold said that unlike typical coding assistants, the app did this work from start to finish.
- comprehensive literature review: Jerome Lecoq at the Allen Institute created a computational review template. It included about 20 custom skills for long-term reviews. Subagents read thousands of papers in the evidence status database. The pipeline then wrote each section using actor-critic agent pairs. Such reviews once took his team up to two years. He now has about 10 reviews, many of which are over 100 pages long.
- genomic epidemiology:Stephen Francis studies the molecular epidemiology of glioma at UCSF. Cloud Science ran germline workups at about one-tenth the rate last time. His group independently confirmed the results.
comparison table
| Dimensions | cloud science | General AI Assistant | cloud code |
|---|---|---|---|
| primary use | scientific research workflow | Q&A and drafting | software development |
| runs the actual pipeline | yes, from end to end | No | yes, code-centric |
| scientific database access | 60+ databases and skills | No | No |
| calculation management | local, hpc (ssh), model | No | local terminal |
| Reproduction/Origin | Complete record of each artwork | No | git history |
| Quote and Number Checking | review agent | No | No |
| native scientific exponent | protein, track, molecule | No | No |
| underlying model | Existing Cloud Model | Existing Cloud Model | Existing Cloud Model |
Cloud science expansion
Cloud Science is an app, so there is no separate inference API. You expand on this through connectors and skills, which persist throughout the session.
You connect a lab tool through a Model Context Protocol (MCP) connector. This is the standard MCP client configuration format:
{
"mcpServers": {
"lab-eln": {
"command": "npx",
"args": ["-y", "@lab/eln-mcp-server"],
"env": { "ELN_API_KEY": "REPLACE_ME" }
}
}
}You save the existing pipeline as a reusable skill. A skill is a folder that contains a SKILL.md file:
---
name: rnaseq-qc
description: Run the lab's standard RNA-seq quality-control pipeline on a FASTQ directory.
---
# RNA-seq QC
1. Run `pipelines/qc.sh <fastq_dir>`.
2. Summarize the per-sample metrics.
3. Flag any sample below the QC threshold.Future sessions inherit these connectors and skills automatically. So you keep your valid devices and data, while the cloud organizes them.
key takeaways
- Cloud Science is a beta app for macOS and Linux; It runs on Anthropic’s existing cloud model.
- A coordinating agent assigns the work, while a separate reviewing agent checks the quotes, numbers, and statistics.
- Each figure comes with its own exact code, surroundings, description and full message history.
- Compute runs locally, on HPC over SSH, or on models, scaling from a single GPU to hundreds.
- It comes with 60+ databases and NVIDIA BioNeMo skills (Evo 2, Boltz-2, OpenFold3) for life sciences.
check it out Technical details here. Also, feel free to follow us Twitter And don’t forget to join us 150k+ml subreddit and subscribe our newsletter. wait! Are you on Telegram? Now you can also connect with us on Telegram.
Do you need to partner with us to promote your GitHub repo or Hugging Face page or product release or webinar, etc? join us
Michael Sutter is a data science professional and holds a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michael excels in transforming complex datasets into actionable insights.
