Claude Science Beta: An AI Workbench for Reproducible Research

Anthropic launched Claude Science in public beta on June 30, 2026: a desktop research app, not a model. It runs analyses, queries scientific databases, and traces every step from data wrangling to publication, attaching the code, environment, and conversation history to each artifact it produces. As of September 2026 it runs on macOS, Windows, and Linux for Pro, Max, Team, and Enterprise plans, and it is local-first: your projects, artifacts, and raw data stay on the machine where it is installed.

The framing matters: this is a workbench wrapped around the model, not another chat surface. For how that fits Anthropic's broader push to turn Claude into a place where work happens rather than a box you paste into, see our coverage of Claude Opus 4.7 and the Claude design launch. Academic and nonprofit labs can get the app through the discounted Claude Team plan for scientists. For the team angle, our internal AI workspaces playbook covers how groups stand up shared, governed AI environments.

Key Takeaways

• It is an app, not a model. Claude Science is a research environment that orchestrates analysis, database queries, and compute. It uses the same Claude models your plan includes.
• macOS, Windows, and Linux. Windows 11 x64 support shipped in version 0.1.47 on September 10, 2026; the app launched on macOS and Linux only. The feature is in beta and under active development.
• Plan gating. On for Pro and Max with no admin action. Off by default for Team and Enterprise until an Owner or Primary Owner turns it on in Organization settings. Not available on the Free plan.
• Your infrastructure, your data. Projects, artifacts, and history live on your computer; Anthropic does not sync them to your account. Prompts and responses still go to Anthropic under its standard retention.
• Reproducibility is the headline. Figures, tables, and notebooks carry attached code, environment, and conversation provenance, and a built-in reviewer flags claims that do not match the execution record.
• Compute you already own. Jobs run on a workstation or HPC login node over SSH (SLURM via sbatch), or on a Modal account that bills you directly; the app connects to 60+ scientific databases.
• Usage comes out of your normal quota. Claude Science counts against each member's standard weekly limit and uses the same seat as claude.ai.

What Claude Science Is

Claude Science is, in Anthropic's words, "your research partner for rigorous science" that "runs analyses, searches databases, and traces every step from data wrangling to publication." The distinction worth internalizing up front: this is a beta app, not a model. Anthropic's product page says it "uses the same Claude models your plan includes. What's new is everything around them: the scientific tools, database connections, and compute integrations that let Claude run full analyses on your own infrastructure."

That changes the unit of work. Instead of pasting context into a chat and copying results back out, the analysis, the data it touched, the compute it ran on, and the conversation that drove it live in one place and stay reconstructable afterward. Claude writes and runs Python, R, and shell commands (PowerShell on Windows) in a persistent kernel that keeps variables, dataframes, and loaded models in memory across the steps of a session; the kernel ends after about 30 minutes idle or when the session ends.

Availability and Plan Gating

As of September 2026, Claude Science is in beta on three platforms. Windows was added in version 0.1.47 on September 10, 2026; the June 30 launch covered macOS and Linux only.

All three need about 5 GB of free disk space for the runtime and starter environments. Plan access, from Anthropic's admin docs:

On Team and Enterprise plans the switch lives in Organization settings > Claude Science, and only an Owner or Primary Owner can flip it; the Admin role can assign seats but cannot turn the app on. Organizations with HIPAA compliance enabled also start with the app off, and Anthropic says usage is not covered under the Business Associate Agreement. Because it is in beta, admins should review the documentation before a broad rollout.

Active scientists at academic and nonprofit institutions can get the app through the discounted Team plan for scientists, verified through the group's principal investigator; our write-up covers eligibility and pricing.

Installing It

Anthropic's Get started doc lists a downloadable installer for macOS and Windows and a shell script for Linux; as of September 15, 2026 it does not mention Homebrew. On macOS you download the DMG for your chip and double-click it; the app then sets up its runtime and starter Python and R environments and opens in a browser tab. On Windows the signed installer adds Claude Science to the Start menu and opens the app in its own window; the first launch downloads a roughly 150 MB app window engine. On Linux you install bubblewrap and socat, run the install script from claude.ai/install-claude-science.sh, then start claude-science serve.

Two placements go beyond a laptop. The Linux command-line version can run on a cloud VM or lab server, with the app used from your own browser through an SSH tunnel; Anthropic says setup takes about five minutes. On Windows, the Linux version also runs under WSL 2 with Ubuntu 24.04 or later, though the Windows app itself does not need WSL.

Sign-in uses your Claude account; no API key is required. A setup wizard then walks through connectors, skills, which websites Claude may reach, and whether memory is on.

Reproducibility and the Reviewer

The reproducibility story is the most concrete part of the pitch. Artifacts (figures, tables, notebooks) carry full provenance: the exact code, the environment it ran in, a plain-language description of what was done, and the conversation that led there. Anthropic's pitch is that results are reproducible months later, by anyone on your team.

On top of that sits the reviewer, which Anthropic's docs describe as "a built-in verification step that independently re-reads Claude's recent responses, the approved plan, saved artifacts, and the execution record, then checks whether its claims match what ran." It runs automatically after responses and periodically during long work, and you can trigger it with Request review. Its examples include a result reported as computed when nothing ran, a value that contradicts the file it came from, a citation that does not support the claim, and a DOI that resolves to a different article. Since version 0.1.41 (August 27, 2026), findings appear as cards directly under the message they refer to.

Two limits are worth knowing. Auto-review is on by default on Max, Team, and Enterprise plans and off by default on Pro, where you turn it on per session. And the reviewer checks claims against the record; it does not re-run analyses and does not judge whether the method was the right one. Anthropic's own caution: "Verify results before relying on them in research, publication, or downstream decisions," and the app "isn't intended for clinical or diagnostic use."

Compute, Databases, and Job Submission

Claude Science is compute-aware rather than compute-blind. It manages work across a local laptop, Linux boxes, HPC login nodes, and GPU clusters, scaling "from one GPU to hundreds" and writing batch scripts as needed. Two remote paths exist:

• SSH hosts. You add a workstation or HPC login node from your existing ~/.ssh/config; the app authenticates with your key or ssh-agent and installs nothing on the host. SLURM clusters receive jobs via sbatch, jobs survive connection loss, and each job needs approval on a card that shows the command and script. The default job timeout is 30 minutes. Remote jobs run outside the sandbox, as your user on the host.
• Modal. Jobs run on a Modal account you own; Modal bills you directly and Anthropic never sees a payment method. Each job is approved with its machine spec and maximum billable time visible. There is no spend ceiling in the app; the default container timeout is 12 hours (maximum 23), and closing the app does not cancel a running job.

It also connects to 60+ scientific databases so you can query them without learning each interface, and ships pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Native viewing covers proteins, structures, molecules, alignments, genomic tracks, chemical structures, and PDFs. Any pipeline can be saved as a reusable skill, any Model Context Protocol server can be added as a custom connector, and NVIDIA BioNeMo NIM model endpoints (Evo 2, Boltz-2, OpenFold3) connect from Settings > Compute on macOS and Linux, not Windows.

What Stays Local and What Anthropic Receives

Anthropic's data doc calls the app "local-first": conversation history and artifacts are stored on the member's computer and are not synced to the Claude account or other devices, so a second computer starts fresh. What does travel is each prompt and Claude's response, logged under Anthropic's standard retention policy for model traffic, plus product telemetry and redacted error reports that device configuration can turn off. Code and data sent to your own SSH hosts or Modal account go directly there and do not pass through Anthropic.

For Enterprise organizations with the Compliance API enabled, Anthropic also retains session transcripts (in beta) that the compliance team can retrieve; they include prompts, responses, tool calls, and the text of files Claude read, but not extended thinking or the system prompt. Admins can turn SSH hosts, Modal, scientific model endpoints, custom connectors, and memory off for the organization.

When to Reach for Claude Science

Why This Matters

The interesting move here is positioning. Most AI-for-science effort goes into better models; Claude Science instead invests in the environment around the model: provenance, compute management, database connectivity, and a reviewer that catches claims the execution record does not support. That is the part of the research workflow that chat surfaces leave to the user, and it is where reproducibility actually breaks down.

The honest caveat is the one Anthropic states: this is beta, it may change, the reviewer reduces but does not eliminate errors, and on Team and Enterprise it is gated behind an Owner-level switch. Treat it as an early but unusually concrete attempt to make AI-assisted research traceable end to end, not a finished platform.

FAQ

See the structured FAQ in the schema header for question-level details: what Claude Science is, which platforms and plans support it, whether it runs on Windows, where data and compute run, how it supports reproducibility, and how usage is billed.

Sources

• Claude Science product page (platform downloads, FAQ on plans, Windows, the scientists discount), checked September 15, 2026: https://claude.com/science
• Anthropic launch post, "Claude Science, an AI workbench for scientists, is now available" (June 30, 2026), the source for the launch-day macOS and Linux scope and the plan list: https://www.anthropic.com/news/claude-science-ai-workbench
• Claude Science docs, overview and requirements (macOS 13+, Windows 11 x64, Linux x64, 5 GB), checked September 15, 2026: https://claude.com/docs/claude-science/overview
• Claude Science docs, Get started (installers, the Linux install script, WSL 2, sign-in), checked September 15, 2026: https://claude.com/docs/claude-science/get-started
• Claude Science changelog (0.1.47 Windows on September 10, 2026; 0.1.41 reviewer cards on August 27; 0.1.14 launch on June 30), checked September 15, 2026: https://claude.com/docs/claude-science/changelog
• Claude Science docs, The reviewer (auto-review defaults by plan, what it checks), checked September 15, 2026: https://claude.com/docs/claude-science/the-reviewer
• Claude Science docs, Remote compute clusters and Compute providers (SSH hosts, SLURM, Modal billing and timeouts), checked September 15, 2026: https://claude.com/docs/claude-science/remote-compute-clusters and https://claude.com/docs/claude-science/compute-providers
• Claude Science docs, Enable Claude Science (plan table, Owner or Primary Owner role, HIPAA), checked September 15, 2026: https://claude.com/docs/claude-science/enable-claude-science
• Claude Science docs, Monitor usage (counts against the standard weekly quota) and How Claude Science works with your data (local-first, retention, Compliance API), checked September 15, 2026: https://claude.com/docs/claude-science/monitor-usage and https://claude.com/docs/claude-science/how-claude-science-works-with-your-data
• Claude Science docs, Tools and environments (Python, R, shell; persistent kernel), checked September 15, 2026: https://claude.com/docs/claude-science/tools-and-environments
• Anthropic Help Center, "Get started with Claude Science" (last updated August 21, 2026; still lists macOS and Linux only, predating the Windows release): https://support.claude.com/en/articles/16563838-get-started-with-claude-science
• Anthropic (@claudeai) launch announcement; the post body cannot be read by our tooling, so no claim on this page rests on it alone: https://x.com/claudeai/status/2072002740830842899