Claude Science: What It Is, Who Can Use It and Where Your Data Goes

You describe a research task in plain language, and Claude writes and runs Python, R or shell code in a sandbox, reads the folders you grant it and saves results as versioned files with a record of how each was made. It isn't only for biologists: Anthropic also lists chemistry, engineering, mathematics, physics and the social sciences, among others.

Who can use Claude Science, and on which plans

You need a Pro, Max, Team or Enterprise plan and a country where claude.ai is offered. Scientists at academic and nonprofit research institutions can get it through the discounted Claude Team plan for scientists, where it's switched on by default. On every other Team and Enterprise plan, an Owner has to enable it in Organization settings before anyone can sign in. It's installed separately from the Claude desktop app.

What you need to know	As of October 2026
Where it runs	Your computer: macOS 13 or later, Windows 11 x64 or Linux x64 on a glibc-based distribution
Plans	Pro, Max, Team and Enterprise, not Free; the Team plan for scientists at academic and nonprofit institutions
Disk space	About 5 GB free for the runtime and starter environments
Usage limits	A 5-hour limit plus a weekly limit, shared with the rest of your Claude plan
Remote jobs	SSH hosts or your own Modal account, billed by Modal

Your files stay put, but what Claude reads doesn't

This is the bit that catches people out. Files are read and written in place, in the folders you grant, and Anthropic doesn't host or store them. But any file content Claude reads to answer a prompt is sent to Anthropic as part of the conversation. If you sign in on a different computer, your files, artifacts and history won't follow you.

Anthropic also receives web search queries, dictation audio while you dictate, optional safety feedback, custom skills you publish and usage telemetry, which carries no conversation content or file names.

On Team and Enterprise plans the Commercial Terms apply, so Anthropic may not train models on Customer Content, and the organisation keeps its inputs and owns its outputs. On Pro and Max, the Consumer Terms and your own Claude privacy settings apply. Requests to models Anthropic designates as Covered Models are kept for 30 days to support its safety work. Saved memory stays in a local database and isn't synced, though facts recalled for a session go out with it.

Permission cards, the sandbox and auto mode

Claude can't quietly reach for more access. A permission card appears each time it needs something new: reading or writing a folder, running code (once, for this conversation, for this project or globally), reaching a network host, using a connector tool or saved credential, or running a remote job. All code runs in an operating-system sandbox that can only touch the workspace and the folders you've granted. Network access is denied by default through a local proxy, which lets through package managers, the databases behind the Featured connectors and hosts you've approved.

Auto mode lets Claude run code and call connectors without asking at every step, while a safety review blocks anything unsafe, such as data exfiltration or prompt injection. Where it's offered, it's on by default for Pro, Max and Team and off by default for Enterprise. It still asks before network access, files outside the workspace, remote compute and package tools.

There's also a background reviewer that checks Claude's claims against the work that was run. It checks the execution record rather than re-running anything, and Anthropic says it reduces errors but doesn't eliminate them. It's a research tool, not meant for clinical or diagnostic use.

Databases, artifacts and the provenance record

Featured connectors cover public research databases, mostly in the life sciences, such as Ensembl, UniProt, gnomAD, ClinVar, PDB, AlphaFold, GEO, PubChem and openFDA, plus OpenAlex and arXiv for literature. They're on by default and can be switched off one by one. PubMed, Clinical Trials, ChEMBL and bioRxiv sit in the directory instead. If connectors are new to you, my guide to how Claude connectors work covers the idea.

An artifact is a file Claude saves into the project: a figure, processed dataset, report or notebook. Every version records its provenance in five tabs, called Messages, Code, Execution Log, Environment and Review. Anthropic's advice is that if the Code tab and the Execution Log disagree, trust the log. Deleting a session keeps its artifacts, but deleting a project deletes everything in it.

Remote compute: the bill that keeps running

Claude can run jobs on a workstation or Slurm cluster over SSH, or on cloud GPUs through your own Modal account. Anthropic doesn't provide or bill compute. Modal bills you directly, per second, and closing the app doesn't cancel a running Modal job.

There's no spend ceiling on Modal jobs from inside Claude Science. Each job is approved on its own, with the machine spec and time limit shown, and up to 10 jobs can run at once by default. The default container timeout is 12 hours, with a maximum of 23.

Usage limits and what costs extra

Claude Science counts toward the same limits as the rest of your plan, including Claude Code and Cowork. Plans with usage limits have one that resets every 5 hours plus a weekly limit. If you hit a limit with no usage credits or extra usage, and the reset is more than 8 hours away, the session pauses for up to 8 hours and then stops, with your conversation and artifacts kept.

You only pay more than your plan price for usage credits or extra usage, usage-based Enterprise spend, or services you connect with your own account, like Modal.

Installing Claude Science, step by step

1. Check you're on Pro, Max, Team or Enterprise. On Team or Enterprise, ask an Owner to enable Claude Science first, unless you're on the Team plan for scientists.
2. Download the installer from the Claude Science product page on macOS or Windows, or run the one-line Linux install from Anthropic's get-started guide.
3. Let the first launch finish. It sets up the runtime and starter Python and R environments, which takes a few minutes, and Windows also downloads a window engine of about 150 MB.
4. Sign in with your Claude account. You don't need an API key.
5. Work through the setup wizard: connectors and skills, which websites Claude can reach, and whether memory is on.
6. Open the Example project or create your own, point Claude at a folder, then approve the folder card and the code card.
7. Open an artifact's provenance tabs to see the code and log behind a result.

Linux needs socat, bubblewrap 0.8.0 or later and unprivileged user namespaces. Ubuntu 24.04's repositories carry a new enough bubblewrap, but Ubuntu 22.04's don't.

Controls for Team and Enterprise owners

Admin controls live under Organization settings, then Claude Science. Owners can switch Featured connectors and skills on or off, add custom connectors and custom skills, set the network allowlist, point the app at an organisation package mirror for conda and PyPI, and control access to previously saved work. There's no organisation setting for web search, dictation or telemetry, though device configuration can switch telemetry off. Usage-based Enterprise plans can enable US-only inference. If your team builds its own skills, my explainer on what Claude skills are is a good starting point.

My verdict on Claude Science

If your research team already pays for Team or Enterprise, I'd trial it, because there's nothing extra to buy and it shares the limits you've already got. Before anyone starts, I'd have the Owner set the network allowlist and decide who's allowed to connect Modal. I'd also tell every user plainly that anything Claude reads goes to Anthropic, and that Pro and Max accounts sit under consumer terms rather than commercial ones. And if anyone runs Modal jobs, make sure each one is stopped when it's finished, because closing the app won't do it for you.

Frequently asked questions

Is Claude Science free?

It's included in the Pro, Max, Team and Enterprise plans with nothing separate to buy, but it isn't available on the Free plan. You only pay extra for usage credits, extra usage, usage-based Enterprise spend or services like Modal.

Does Anthropic store my Claude Science files?

No. Files are read and written in place on your computer, but file content Claude reads to answer a prompt is sent to Anthropic as part of the conversation.

Which computers does Claude Science run on?

macOS 13 or later, Windows 11 x64 and Linux x64 on a glibc-based distribution, with about 5 GB of free disk space.

Can Claude Science be used for clinical decisions?

No. Anthropic describes it as a research tool not intended for clinical or diagnostic use, and says its background reviewer reduces errors but doesn't eliminate them.