Claude Science capabilities

Claude Science ships with 60+ pre-configured, read-only connectors to major scientific databases and modeling tools spanning genomics, structural biology, and cheminformatics, plus a directory of skills for domain-specific workflows (e.g. structure prediction, protein design, single-cell analysis).

Claude Science is architected as a local-first desktop app: conversation history and artifacts live only on the member's device, with no Anthropic-hosted session store. Only the model prompt/response for each step is sent to Anthropic; large or sensitive datasets and remote-compute traffic never pass through Anthropic's servers.

Claude Science is bundled into existing paid Claude plans rather than sold as a standalone product: it requires a Pro, Max, Team, or Enterprise account and is not available on the Free plan. There is no separate Claude Science subscription or usage-based add-on fee at launch.

Claude Science can hand off heavy analysis to remote infrastructure — personal Linux boxes, on-premise HPC clusters, or on-demand GPU providers via Modal — letting a research session scale from a single laptop to hundreds of GPUs without leaving the chat interface.

Claude Science is a desktop research app where a generalist coordinating agent orchestrates 60+ curated skills and connectors to run scientific analyses, draft manuscripts, and manage a persistent research workspace. Sessions organize into projects with shared folder permissions and custom instructions, and work can be split into parallel tracks via delegation.

Claude Science runs Python, R, and shell code in a local sandboxed environment on the user's machine, keeping persistent kernels alive across a session so variables, dataframes, and loaded models carry between analysis steps instead of resetting each turn.

Claude Science is off by default for Team and Enterprise organizations; an Owner must turn it on under Organization settings > Capabilities before any member can use it. Once enabled, admins get identity, connector, and usage-monitoring controls, though several enterprise controls common to other Claude products are still beta-limited for Claude Science specifically.

Every result Claude Science produces (figures, notebooks, structures, manuscripts) saves as a versioned artifact with a full provenance record — the exact code, environment, and conversation that produced it. A background reviewer agent cross-checks Claude's claims against that record to catch citation errors, unverifiable numbers, and figures that don't match their underlying code.