Move your team from Claude to Claude Science.
Each person gets a short interview, then a personalized step-by-step guide showing exactly where their saved work, prompts, and projects land in Claude Science — and the few things that need a manual step, walked through. You send one invite; nobody becomes the help desk, and nothing in Claude is changed or cancelled. Want to see where you’re paying for both first? Run the free audit.
Claude Science is included at no extra charge on the same plans that already unlock it per the base Claude pricing table (Pro and above; explicitly listed alongside Research, Claude Code, Claude Cowork, and Claude Design as a Pro-tier inclusion), but plan access alone is not enough for Team/Enterprise members. An org Owner must separately enable it under Organization settings > Capabilities before any member — regardless of role — can use it; only after that does the org's existing SSO/SCIM/role-management apparatus extend to Claude Science, with some carve-outs (no audit logging or Compliance API export for Claude Science yet, since its data is local-first). Admins planning a rollout should treat 'plan supports it' and 'org has switched it on' as two separate gates, and confirm the enterprise controls they rely on for other Claude products (audit logs, Compliance API, custom retention) don't silently apply to Claude Science.
- Warning: Enabling Claude Science is all-or-nothing per org at launch — there is no per-role opt-in, so an Owner cannot restrict it to only certain teams within the same Organization settings toggle.
- Warning: Because Claude Science data is local-first (lives on the member's device), several enterprise controls that work for other Claude products don't apply: no Custom Data Retention, no Org Data Export, and revoking a member's access does not delete their local data — that requires separate MDM/EDR handling, unlike other Claude data governed centrally.
- Warning: IP allowlisting for Claude Science is only 'partial' per the admin docs, and data-processing-geography controls explicitly exclude remote compute traffic (SSH/HPC/Modal jobs) — don't assume full parity with other Claude products' geography/compliance controls.
- Warning: A discounted Team plan exists for academic/nonprofit research institutions, which may change the cost calculus for a lab-scale rollout versus a standard commercial Team plan.
Claude's standard code execution tool runs Python/Bash in a hosted, ephemeral sandboxed container inside a chat, billed by container-time on the API and included from the Free plan in claude.ai. Claude Science's sandboxed compute environment is the local, persistent-kernel evolution of the same idea: it runs on the user's own machine (not a hosted container), keeps Python/R kernels alive across an entire session so state survives between steps, and adds explicit per-approval gating for new folder access, network hosts, and remote jobs. A researcher whose workflow depends on chat-based code execution resetting cleanly each session should note Claude Science instead accumulates state deliberately — better for iterative analysis, but it means workspace folders need active management rather than disappearing automatically.
- Warning: Standard Claude code execution requires no local install and runs on Free/Pro/Max/Team/Enterprise across web, desktop, and mobile; Claude Science's sandbox requires a local desktop install (~5GB disk, OS-level prerequisites: bubblewrap/socat/user namespaces on Linux, macOS 13+) and has no mobile or web surface at all.
- Warning: Standard code execution has no R support (Python/Bash only); Claude Science adds R kernels, a capability difference in Claude Science's favor rather than a migration blocker.
- Warning: Regular code execution's container is stateless across chats and network-isolated unless paired with web_search/web_fetch tools; Claude Science's sandbox can access the network and remote hosts with per-item user approval, which changes the security posture a team needs to review before adopting it.
- Warning: Claude Science has no Windows-native build (WSL only, and not yet fully available per its own docs), whereas standard Claude code execution works on Windows via the web/desktop app — this is a hard platform gap, not just a rough edge.
Claude's general Connectors let a user or org attach MCP integrations for productivity SaaS (Google Drive, Gmail, Slack, GitHub, Notion, etc.) via the '+' menu, enabled per-conversation. Claude Science replaces that general-purpose set with 60+ connectors purpose-built for science — genomics (Ensembl, UniProt, ClinVar, GWAS Catalog), structural biology (PDB, AlphaFold, STRING), and cheminformatics (PubChem, ChEBI, BindingDB) — enabled once in Settings > Connectors with no per-conversation re-approval and no auth required for the featured set. A researcher already comfortable enabling Claude Connectors should expect Claude Science's featured connectors to work more automatically (no OAuth flow), but custom/non-featured scientific data sources still follow the same MCP custom-connector approval model (Once/conversation/project/global) as regular Claude.
- Warning: Regular Claude Connectors are general business-productivity integrations; none of the standard connector list (Drive, Slack, GitHub, etc.) is replaced or included by default in Claude Science's featured set — the two lists serve different purposes and a Team/Enterprise admin enabling one does not enable the other.
- Warning: Claude Science's featured connectors are strictly read-only, unlike some standard Claude connectors which can take actions (e.g., interactive 'MCP Apps' with write behavior) — a workflow that relies on a connector writing back to a source has no equivalent among Claude Science's featured set.
- Warning: Standard Claude connector restrictions (Team/Enterprise Owner enabling org-wide, blocking outside-org connections) are described as still maturing for Claude Science specifically — its admin doc notes org-published Directory connectors can be role-restricted, but locally-added custom connectors and plugins are not restricted, a gap versus the general Claude connector admin model.
Claude's Research mode already runs a lead-agent-plus-sub-agents pattern (Opus orchestrator, Sonnet sub-agents) inside claude.ai to produce a cited report against the web and Connectors. Claude Science's coordinating agent extends the same multi-agent idea into a persistent desktop research workspace: sessions carry live kernels and a workspace folder rather than ending when the report is delivered, projects add shared folder permissions and custom instructions across sessions, and delegation lets a user fork a session into parallel tracks to compare analytical approaches — closer to spinning up multiple Research runs but without losing shared context. A Claude Pro/Max/Team/Enterprise user already using Research for literature synthesis can move deeper technical/computational research work into Claude Science while keeping Research in claude.ai for pure web-citation tasks; both are available on the same paid plans (Pro and above) with no additional purchase.
- Warning: Claude Science is a separate desktop app (beta, macOS Apple Silicon/Intel and Linux x64 only, no Windows-native or web build) — it is not simply a mode switch inside claude.ai the way Research is; users must install a distinct application.
- Warning: Research's ~15x token cost is against the standard chat usage budget; Claude Science sessions (including the optional Reviewer) draw from the same shared 5-hour/weekly limit as Claude Code and Claude Cowork, so heavy Claude Science use can compete with Research/Code/Cowork usage in ways a claude.ai-only user wouldn't have budgeted for.
- Warning: Research's Connectors (Google Drive, Gmail, SharePoint) carry over conceptually but Claude Science's default toolset is a different, curated set of 60+ scientific-database connectors — general productivity Connectors aren't the focus and may need separate enabling.
Claude Projects group conversations under shared custom instructions and uploaded knowledge files, giving a stable context for recurring work. Claude Science's artifact provenance and Reviewer system builds a similar durable record but centers it on outputs rather than instructions: every figure, notebook, structure, or manuscript is saved as a versioned artifact carrying the exact code, environment, and conversation that produced it, and a background Reviewer agent automatically cross-checks Claude's claims against that record. A user relying on a Claude Project's knowledge files to keep a research thread consistent should expect Claude Science's per-artifact provenance to answer a narrower but more rigorous question — not 'what context informed this conversation' but 'exactly what execution produced this specific number or figure' — and to get an automated correctness pass Projects don't offer at all.
- Warning: Claude Projects scope memory and knowledge at the project level across many conversations; Claude Science provenance is scoped per-artifact within a session/project workspace and is about output traceability, not shared instructions — the two aren't interchangeable even though both use the word 'project'.
- Warning: The Reviewer is automatic on Max/Team/Enterprise but manual-only ('Request review') on Pro; a Pro-plan user migrating from Projects should not assume the correctness check runs unprompted the way Project-level context does.
- Warning: The Reviewer explicitly checks claims against the execution record and cannot judge whether a chosen scientific method was appropriate — it is not equivalent to a human collaborator reviewing a Project's shared knowledge for correctness.
- Regular Claude Projects are available from Free (capped at 5 projects); Claude Science overall requires Pro/Max/Team/Enterprise, so a Free-tier Projects user has no path to the Reviewer/provenance capability without upgrading.