Anthropic launches Claude Science: an AI workbench for scientists, with 60+ built-in research skills
- Anthropic has launched Claude Science, an AI workbench app for scientists, now in open beta for Pro, Max, Team, and Enterprise users, usable locally on macOS/Linux or remotely via SSH/HPC login nodes.
- The app comes with 60+ pre-configured skills and connectors covering genomics, single-cell biology, proteomics, structural biology, and cheminformatics, backed by connections to hundreds of specialized data sources (UniProt, PDB, Ensembl, etc.) as well as journals and preprint resources.
- It can autonomously draft computational jobs and, with the user's consent, submit them to the user's own HPC cluster or Modal cloud GPUs, scaling analysis from a single GPU to hundreds, while raw data always stays on the user's own systems.
- A built-in reviewer agent checks throughout whether citations in generated content are real, whether numbers trace back to the computation, and whether figures match the code that produced them, auto-correcting issues it finds.
- Real-world cases already exist: an Allen Institute researcher produced roughly ten literature reviews (several over 100 pages) that used to take two years each; a UCSF team cut a germline variant analysis pipeline down to one-tenth the time, independently verified by the lab.
Scientists now have their own AI workbench
Anthropic recently launched Claude Science, an AI workbench app that brings together the tools, databases, and compute resources scientists rely on into a single environment, open for beta to Pro, Max, Team, and Enterprise users.
It's an app installed on your own computer or server: you ask a scientific question in plain language to an AI, and it mobilizes dozens of specialized tools to query data, run analyses, generate charts, and draft manuscripts, with every step traceable back to how it was produced. You can use it locally (macOS/Linux) like a Jupyter Notebook, or on a remote machine via SSH or an HPC login node.
How frustrating scientists' day-to-day really is
There's a huge amount of tedious work baked into science. Researchers have to hop between dozens of databases, each with its own schema; the file formats they run into often need dedicated pipelines and viewers; and the toolset is a long list — PubMed, Jupyter, R, cluster terminals — switched one after another.
Just getting these tools to work together, and letting data flow between them, eats up a huge amount of a researcher's energy. What Claude Science wants to do is pull these scattered pieces into one environment, so people can go from literature search to manuscript inside it.
One coordinator agent, a team of experts working behind it
You're talking to a general-purpose coordinator agent. It holds those 60+ pre-configured skills and connectors, and can rally domain-specialist sub-agents, as well as custom agents you've built yourself. They go off separately to query data, run analyses, and produce results.
Think of it like an app store plus shortcuts on your phone: whatever database or piece of software you need to operate, you equip the coordinator agent with the matching "skill pack," and it knows how to call that tool. Connectors, meanwhile, hook up the tools your lab already uses.
The key piece is that there's a dedicated error-catcher at the end of the chain: the reviewer agent. It watches the output of the other agents, checks it item by item, and fixes problems itself when it finds them.
ask in plain language
Auto-corrected
This reviewer agent works like an always-on peer reviewer: it specifically checks whether citations actually have a source, whether the numbers given can be traced back to the original computation, and whether figures match the code that generated them. If it finds an error, it fixes it itself rather than leaving the problem for you. This goes straight at the old complaint that "AI-generated content confidently makes things up."
One agent generates content — the "actor" — while another is dedicated to checking accuracy and citation reliability — the "critic" — and the two keep each other in check. It's like one reporter writing a story while another, a dedicated fact-checking editor, verifies every sentence; neither one vouches for the other.
Every figure it generates traces back to its code
Science itself relies heavily on visuals, so when Claude Science produces figures and manuscripts, it hands over the code that generated them alongside. It can also natively render science-specific visualization formats — 3D protein structures, genome browser tracks, chemical structures — without needing a separate viewer.
When it produces a figure, it hands over along with it: the exact code and runtime environment that produced it, a one-line explanation of how it was derived, and the full conversation record. That means months later, you can still tell what went in, how the result was validated, and how to reproduce it.
- The code that produced it
- Runtime environment
- A one-line explanation
- Full conversation record
- See every input clearly
- Verify anytime
- Reproduce it months later
Editing a figure doesn't require touching code yourself, either. Just tell it in plain language, "remove the gridlines" or "switch the y-axis to log scale," and it will edit the code it wrote and regenerate the figure.
AI operates your supercomputer directly — data never moves
Big analyses are a hassle: folding a protein, running a genomics pipeline over a massive dataset — researchers often have to stop working on the actual scientific question to configure a compute job, wait in a cluster queue, watch to see if it succeeded, then pull the results back. Claude Science takes over that entire loop.
It first drafts a plan and asks you before touching any new resources — you can review or even revoke any decision. Once approved, it writes the job and submits it to the compute your lab already uses: your own HPC cluster over SSH, or on-demand cloud GPUs through your Modal account. It scales from a single GPU up to hundreds.
reviewable / revocable
／ Modal GPU
Context Returns
The whole process runs on your lab's own infrastructure — your laptop, a Linux machine, or an HPC login node. So large, sensitive datasets never have to leave the systems they're already sitting on — only the minimal context each analysis step needs gets passed to Claude. Compute gets outsourced to AI scheduling; the raw data doesn't move.
Because these agents work within a session that keeps context in memory, even massive datasets only need to be loaded once. While a job runs, the reviewer agent checks the output in parallel, catching bad citations, untraceable numbers, and figures that don't match the code, self-correcting as it goes.
Partway through a task, you can duplicate the session into a parallel branch, try a different approach in each without them interfering with one another, while the original conversation thread stays intact. It's like saving the same document as two separate versions to edit independently — if one gets messed up, the original is untouched.
Ready out of the box: databases and specialized models already connected
Scientific knowledge is scattered across hundreds of specialized sources. In biology alone, relevant data may be spread across UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO — each with its own structure and query language — plus journals, preprint servers, and domain-specific open-source models. Ask a question in plain language, and a specialist agent will query and synthesize across these sources, saving you from having to navigate each one yourself.
It also connects to NVIDIA's BioNeMo Agent Toolkit, natively linking to life-science models and libraries in BioNeMo, including Evo 2, Boltz-2, and OpenFold3. And the models, datasets, and pipelines scientists already trust can be brought in too: any pipeline can be saved as a reusable skill, any commonly used tool can be hooked up via a connector, and future sessions automatically inherit them. You don't have to give up the toolchain you already trust just to use AI.
What three labs have already done with it
Over the past several months, researchers testing the beta have used it for single-cell RNA-seq analysis, CRISPR screen design, protein structure prediction, cheminformatics, and more. Three cases best show what it looks like in practice.
| Lab | What they did with it | Quantified result |
|---|---|---|
| Manifold Bio | End-to-end screening of tissue-targeting drug candidates, evaluating surface expression, in vivo trafficking, and safety for each target, ranked against criteria learned from their own proprietary data | Ran the full pipeline in one pass; the key difference from a general-purpose coding assistant is that it can find the right data itself and make judgment calls informed by past project experience |
| Allen Institute Neuroscientist Jérôme Lecoq | Built a multi-agent "computational review template" with about 20 custom skills; sub-agents read thousands of papers, extract core arguments and key quantitative findings into an evidence base, then draft the review section by section, with each section handled by a dedicated sub-agent using actor-critic pairing to check as it writes | A review that previously took up to two years to write has now produced about 10, several over 100 pages, with citations checked by the reviewer agent |
| UCSF Brain Tumor Center Associate Professor of Epidemiology Stephen Francis | Researching the molecular epidemiology of glioma: how thousands of small-effect germline variants combine to shape individual susceptibility, running a comprehensive germline analysis across multiple methods | Time cut to about one-tenth; the team independently reviewed the results and confirmed they were both fast and reliable |
Who can use it now, and how
The Claude Science app is now in beta on macOS and Linux for Pro, Max, Team, and Enterprise users. Team and Enterprise users need an admin to enable it. Anthropic says releasing early is meant to let scientists start working on real problems now, and feed back what needs refining.
There's also a discounted Team plan seat tier for active labs at academic institutions and nonprofit research organizations.
There's also a science-program grant (click for details)
Anthropic will support up to 50 Claude Science "AI for Science" projects, each with up to $30,000 in credits; Modal is separately providing up to $2,000 in compute for selected projects. Priority goes to biology and biomedical research. Applications are open through July 15, 2026, with notifications by July 31, and the project period runs September 1 through December 1, 2026.
Every output comes with an auditable record of how it was generated, so you can verify and reproduce the results. Anthropic, "Claude Science, an AI workbench for scientists"
Science goes from a desk full of disconnected tools → one AI workbench that does it all, with a built-in error-catcher
Anthropic launches Claude Science: scientists ask a question in plain language, and AI queries data, operates supercomputers, and drafts manuscripts for them — while checking its own citations and numbers, and never letting raw data leave the local system. One page, with a figure.
↓ One page to read it all · one figure animates
Claude Science is an app Anthropic just launched, installed on a scientist's own computer or server. Ask a scientific question in plain language, and it mobilizes dozens of specialized tools to query data, run analyses, generate charts, and draft manuscripts for you. Why is it needed? First look at how fragmented scientists' days currently are.
✘ But they each sit in their own corner, speaking their own language, and the human has to act as the go-between
Dozens of databases each have their own data schema, and many file formats need dedicated pipelines and viewers; researchers spend all day switching between these tools, and just getting them to work together and let data flow smoothly eats up a huge amount of energy.
The change is this: those scattered tools, compute resources, and manual checking have all been folded into one conversational AI workbench. The AI drafts its own compute jobs, and once you consent, submits them to your lab's own supercomputer (HPC, high-performance computing cluster) or on-demand cloud GPUs (chips dedicated to running large-scale computation).
- Dozens of databases, PubMed, Jupyter, R, cluster terminals, all switched by hand
- Compute had to be configured, queued, watched for success, then results pulled back manually
- Citations and numbers the AI produced had to be checked one by one, by hand
- A coordinator AI orchestrates everything; one sentence sets multiple agents to work
- The AI drafts the compute job, and once you approve, submits it to the supercomputer automatically
- Another AI checks citations, numbers, and figures as the job runs, auto-fixing errors it finds
Handing the work to AI immediately raises two old questions: will it misuse compute or leak sensitive data out? Could the citations and numbers it generates be made up? The figure below shows how it addresses both.
You only ever talk to one "coordinator agent" (the overall dispatcher AI), which rallies a team of specialists behind the scenes; at the end stands a "reviewer agent" (an AI dedicated to catching other AIs' mistakes), checking item by item and fixing errors itself.
The official framing is something like "10x faster," which doesn't mean much on its own. Converted into how long the same task took under the old way versus the new one:
why am I still exhausted?
- × Dozens of DBs
each its own dialect - × PubMed·R·Jupyter
switching back and forth - × Compute jobs —
queue, watch, wait
raise glioma risk?
will my data leak out?
only the essentials get passed
it writes be made up?
fix it the moment it's wrong
ninety percent less time
is enough to ask more questions