Anthropic Built a Research Agent for Scientists, Hired a Nobel Laureate, and Made It Run on Your Own Machine
TL;DR
On June 30, Anthropic shipped Claude Science, an AI workbench that turns a research project into a multi-agent job: a coordinating agent with more than 60 curated skills spawns specialists, a reviewer agent checks the citations and the arithmetic, and everything comes out as figures and manuscripts bundled with the code that made them. The twist for builders is where it runs. Not on Anthropic's GPUs, but on your laptop, your Linux box, or your HPC login node, so sensitive datasets never leave the building. It is in beta for Pro, Max, Team, and Enterprise, no enterprise vetting required. And in the same stretch, Anthropic hired the Nobel laureate behind AlphaFold and announced it will start hunting its own drugs.
What Claude Science Actually Is
Strip the launch-event gloss and Claude Science is an orchestration harness pointed at science instead of software. You give it a high-level goal, it plans the steps, pulls literature, runs code, and iterates on the output. It ships preconfigured for four domains a bench scientist actually lives in: genomics, single-cell, proteomics, and cheminformatics, with connectors into more than 60 scientific databases so the model is not guessing at data it cannot see.
The structure is a team, not a chatbot. A generalist coordinator holds the plan and delegates to specialist sub-agents, then a separate reviewer agent audits the citations and calculations before anything is presented. If you have ever run a lab, the shape is familiar: a principal investigator hands pieces to grad students, and a skeptical postdoc checks the math before it goes anywhere near a paper. The difference is that here the postdoc never gets tired of double-checking a p-value.
The Part That Matters for You: It Runs on Your Box
Here is the line that should make homelab people sit up. Claude Science is built to run "on your lab's own infrastructure, your laptop, Linux box, or HPC login node." The sensitive dataset stays local; the compute happens where the data already is. A frontier lab shipping software that executes on your hardware instead of its own is either generous or strategic, and it is strategic: it is the only way a hospital or a pharma company with patient data or unpublished results will let an agent near the good stuff.
It runs on Anthropic's flagship Opus model, the same one paying subscribers already get, with no special access tier. One Harvard physicist, quoted by MIT Technology Review, estimated the model is about as capable at executing scientific projects as a second-year graduate student. That is the whole pitch in one sentence: not a Nobel in a box, a competent junior researcher who works nights.
Anthropic Poached a Nobel Laureate
The product did not appear in a vacuum. On June 19, John Jumper announced he was leaving Google DeepMind for Anthropic after nearly nine years. Jumper is not a random hire. He shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for AlphaFold, the deep-learning system that predicted the structures of more than 200 million proteins and rewrote structural biology. When the person who won a Nobel for AI-in-science walks from the lab that built AlphaFold to the lab shipping Claude Science, that is a signal, not a footnote.
Anthropic Is Now Its Own Customer
Anthropic also said it will use Claude Science internally to pursue its own research into drug candidates for neglected diseases. Read that twice: the company selling the research tool is now also running the research. It is the classic platform-versus-first-party tension, arriving on day one. If your agent can do the science, the temptation to keep some of the science for yourself is enormous, and Anthropic is not pretending otherwise.
To seed adoption, Anthropic opened an "AI for Science" program: up to 50 projects receive up to $30,000 in credits each, aimed at academic and nonprofit research. Applications close July 15, awards are announced by July 31, and funded projects run from September 1 to December 1, 2026. There is also a discounted Team plan for universities and nonprofits.
The Caveats, Straight
Keep the excitement calibrated. The "second-year grad student" line is one researcher's estimate, not a benchmark, and executing a project well is not the same as knowing which project is worth doing. The reviewer agent checks citations and calculations, but a model that fabricates a plausible-looking source is exactly the failure mode a citation checker can miss if the fabrication is internally consistent. Beta means macOS and Linux only for now, and Team or Enterprise seats need an admin to switch it on. And the first-party drug program means the platform's incentives and its customers' incentives are not perfectly aligned, which is worth watching as this matures.
Key Takeaways
- Claude Science launched June 30 as a multi-agent research workbench: a coordinator with 60+ skills, specialist sub-agents, and a reviewer agent that checks citations and math.
- It runs on your own hardware, laptop, Linux box, or HPC login node, so sensitive data stays local. That is the real unlock for regulated and unpublished research.
- Availability is broad, not gated: beta on Pro, Max, Team, and Enterprise (macOS and Linux), preconfigured for genomics, single-cell, proteomics, and cheminformatics across 60+ databases.
- Anthropic hired AlphaFold's John Jumper from DeepMind, a 2024 Nobel laureate, telegraphing how seriously it takes AI-for-science.
- Anthropic will run its own drug programs for neglected diseases, making it both the toolmaker and a competitor to its users.
- Grants are open: up to 50 projects, up to $30,000 in credits, applications close July 15, projects run September to December 2026.
Sources: Anthropic: Claude Science, an AI workbench for scientists, MIT Technology Review, The Nobel Prize in Chemistry 2024, Google DeepMind: AlphaFold