Claude Science Launch: Key Takeaways for Biotech
Anthropic released Claude Science, and we’ve begun exploring its potential internally and with customers.
Our initial reaction is that Claude Science is a brilliant example of why AI enablement in biotech needs to be more than tool-by-tool approval.
The question is no longer simply whether scientists should be allowed to use AI assistants as many already are. The more important question is how organizations can safely extend AI from chat-based assistance into real scientific workflows involving code, compute, proprietary data, public databases, scientific artifacts, and, in some cases, regulated or decision-sensitive work.
That is why we believe biotech organizations need minimally viable governance for AI. The goal is not to slow scientists down with a draconian process before they can experiment. It is to create enough structure that teams understand what data they can use, which tools and compute environments are approved, how outputs should be reviewed, and where sensitive or regulated work requires additional care.
This is also where we like to champion a flywheel model for AI enablement as innovation and governance should reinforce each other rather than compete. Teams should start with real scientific use cases, learn where AI creates practical value, and then turn those lessons into reusable guidance, skills, workflows, and controls. As those patterns mature, governance becomes less of a gate and more of an accelerator, giving the next team a safer and clearer path to adopt AI without starting from scratch.
Mantle take: the most valuable version of Claude Science will not be “every scientist improvises.” It will be “scientists use governed, reusable skills, vetted connectors, and plugins that reflect the company’s methods, standards, and risk posture.”
What is Claude Science
Claude Science is Anthropic’s new AI workbench for scientific research. Currently in beta, it is available on macOS and Linux for Claude Pro, Max, Team, and Enterprise users, with Team and Enterprise deployments requiring administrator enablement.
Importantly, Claude Science is not a separate foundation model or a specialized domain specific model around chemistry or biology. It uses the off-the-shelf Claude models included in a user’s plan. However, what is new is the specifically curated research environment:
A life-science-curated workspace that brings together scientific tools
Database connectors
Code execution
Compute orchestration
Reproducible artifacts
Reviewer agents
Domain-specific skills
For biotech companies, Claude Science further differentiates Anthropic as a serious AI enabler for life sciences. More importantly, it continues to shift the conversation from whether organizations should use AI assistants at all to how they should begin using them responsibly — and how they can safely extend AI from chat-based assistance into real scientific workflows involving code, compute, proprietary data, public databases, scientific artifacts, and, in some cases, regulated or decision-sensitive work.
Claude Science is also a strong example of why Mantle believes that biotech organizations need minimally viable governance.
The goal is not to slow scientists down with a heavyweight approval process before they can experiment. It is to put enough structure in place that teams understand what data they can use, which tools and compute environments are approved, how outputs should be reviewed, and where sensitive or regulated work requires additional care.
This connects directly to what we champion, which is a flywheel model for AI enablement. We believe innovation and governance should reinforce each other rather than compete. Teams should be encouraged to start with real scientific use cases, learn where AI creates practical value, and then turn those lessons into reusable guidance, skills, workflows, and controls. As those patterns mature, governance becomes less of a gate and more of an accelerator. It gives the next team a safer and clearer path to adopt AI without starting from scratch.
That is the business-partner role IT and cybersecurity need to play. The work is not to block tools like Claude Science, nor is it to approve them casually. The work is to help the organization create enough trust, structure, and repeatability that scientists can move faster without creating avoidable risk.
The basic idea
Claude Science is designed to let researchers describe scientific tasks in plain language, then have Claude:
Write and run Python, R, and shell code for data analysis and interpretation
Query scientific databases
Manage local or remote compute
Save outputs as versioned artifacts with provenance
Anthropic positions it as a workbench that brings together tasks that otherwise span many disparate solutions, including PubMed, Jupyter, R, terminal sessions, HPC clusters, figures, manuscripts, and domain-specific databases.
The product is especially relevant for computational biology, genomics, single-cell analysis, proteomics, structural biology, cheminformatics, literature review, indication dossiers, scientific figure generation, and reproducible analysis workflows. Anthropic says Claude Science is pre-configured for major life sciences domains and can query more than 60 scientific databases.
Claude Science is designed to let researchers describe scientific tasks in natural, plain language and have Claude help carry the work forward across their tools and environments. Instead of stopping at a chat or cowork response, Claude can:
Write and run Python, R, and shell code for data analysis/interpretation
Query scientific databases
Work across local or remote compute environments
Generate figures and manuscripts
Save outputs as versioned artifacts with provenance
In Anthropic’s framing, the product brings together workflows that often require researchers to move between disparate solutions, including PubMed, Jupyter notebooks, R, terminal sessions, HPC clusters, domain-specific databases, figures, and manuscript drafts.
Anthropic has enabled Claude Science as a pre-configured utility for major life sciences domains and backed by more than 60 scientific databases, with curated skills and connectors intended to make the product useful to scientists on day one rather than requiring each organization to build the entire environment from scratch.
Claude Science should be read as more than a feature release. It suggests that Anthropic is trying to move Claude from a general-purpose assistant into more specialized work environments where professionals can use AI inside the actual workflows of their domain. Claude Code did this for software development by bringing Claude closer to repositories, terminals, debugging, and implementation work. Claude Science appears to be the analogous move for scientific research: bringing Claude closer to literature, code, compute, scientific databases, artifacts, and reproducible analysis.
Endpoints News framed the launch in a similar way, describing Claude Science as part of Anthropic’s broader push into biopharma and as a productized life sciences workbench rather than simply another model release. That context matters because it positions Claude Science not as a science-themed chatbot, but as part of a larger effort to make Claude useful inside the practical, strategic, and computational workflows of drug development.
Dimension 1: Compute options
Claude Science is most interesting because it is not limited to a hosted chat environment. It can work where scientific data and compute already live.
| Compute option | What it means | Best fit | Security / governance implication |
|---|---|---|---|
| Local laptop or desktop | Claude Science can run code locally in a sandboxed analysis environment. Python and R can run in persistent kernels, keeping variables in memory during a session. | Exploratory analysis, literature work, small datasets, figure iteration, and method prototyping. | Data may stay local, but prompts and model responses are still sent to Anthropic. Endpoint controls matter because artifacts and conversations are stored on the user’s device. |
| Lab Linux box | The app can run where data already lives, including a Linux machine. | Shared scientific workstations, controlled lab compute, and local GPU work. | This can be a strong middle ground for research compute, but requires device hardening, access control, storage governance, and monitoring. |
| GPU-enabled local Linux machine | If a Linux machine has GPUs, Claude Science can make them available to code after GPU access is enabled. | Structural biology, model inference, and GPU-heavy scientific workloads. | Treat this as elevated execution. Approval, logging, package governance, and machine isolation become more important. |
| Remote workstation or HPC cluster over SSH | Claude Science can connect to machines reachable over SSH, including lab workstations and HPC login nodes, using the user’s existing SSH configuration and keys. | Larger genomics workflows, structural biology workloads, and institutional compute environments. | Remote jobs run as the user on the destination host and can access what that account can read or write. This requires clear identity, access, storage, and monitoring controls. |
| Modal cloud compute | Claude Science can run jobs through a Modal account owned and controlled by the user or organization. | On-demand GPU or memory-heavy workloads when internal compute is insufficient. | Useful for elasticity, but requires cloud spend governance, approved data boundaries, and clarity on where outputs are stored. |
Dimension 2: Skills and connectors
Claude Science uses two key extension concepts: connectors and skills.
Connectors give Claude access to external data sources during an analysis. Skills are instructions Claude loads when relevant, telling it how to run a method, which tools to use, and what to verify. Anthropic says both are managed in Settings and apply across projects.
Featured scientific connectors
Claude Science includes featured read-only connectors to many public life sciences databases. Examples include Ensembl, UCSC, MyGene, UniProt, GO, Reactome, ClinVar, dbSNP, gnomAD, GWAS Catalog, GTEx, ENCODE, PDB, AlphaFold, EMDB, GEO, ArrayExpress, PRIDE, PubChem, ChEBI, BindingDB, openFDA, OpenAlex, arXiv, Grants.gov, and others. Anthropic also lists directory connectors such as PubMed, Clinical Trials, ChEMBL, and bioRxiv.
Featured science skills
Anthropic lists science skills for literature review, indication dossiers, AlphaFold2, Boltz-2, Chai-1, ESMFold2, OpenFold3, ProteinMPNN, DiffDock, ESM-2, Evo 2, Borzoi, scGPT, and scvi-tools. Users can also create skills from scratch, import from GitHub, upload a skill, or ask Claude to distill a workflow from an existing session into a reusable skill.
Why skills matter
Skills are where organizational know-how can become reusable. For a biotech, this could include:
A company-standard literature review workflow.
A target assessment template.
A bioinformatics QC workflow.
A reproducible figure-generation style.
A clinical or regulatory evidence synthesis workflow.
A “how we evaluate a disease area” indication dossier process.
A preferred method for summarizing public, proprietary, and partner data.
The strategic opportunity is to turn scattered individual prompting into a shared scientific operating model.
Mantle take: the most valuable version of Claude Science will not be “every scientist improvises.” It will be “scientists use governed, reusable skills that reflect the company’s methods, standards, and risk posture.”
Dimension 3: Data security and governance
Claude Science is local-first, but it is not data-free from Anthropic. For biotech organizations, this creates a split control model across the user’s device, Anthropic’s model traffic, approved compute environments, and any connectors used in the workflow.
Anthropic states that Claude Science stores conversation history and artifacts on the member’s device rather than in an Anthropic-hosted session store. When Claude Science calls the model, however, prompts and Claude’s responses are sent to Anthropic and handled under Anthropic’s standard model-traffic retention and Trust & Safety policies. Anthropic also notes that some server-side enterprise controls, including Custom Data Retention, Org Data Export, and the Compliance API, do not extend to Claude Science conversations and artifacts stored locally on user devices.
For biotech companies, this means Claude Science should not be reviewed like a typical SaaS application or treated like a standard chatbot. It touches local files, scientific data, code execution, model traffic, compute environments, connectors, credentials, and research outputs. The governance question is not only whether Anthropic is following good due care and diligence, it is also that you, as the end-customer, also need to understand how the full workflow behaves across devices, data sources, compute environments, and users.
The practical concerns for management and leadership teams are as follows:
Business data protection: Scientists should be able to use powerful tools without accidentally exposing sensitive IP, personal data, clinical data, partner data, regulated records, or board-sensitive information.
Validation of AI-assisted work: The organization should have confidence that AI-generated or AI-assisted outputs are reviewed before they influence scientific, operational, or business decisions.
Transparency and traceability: Teams should be able to explain what data was used, where it went, who had access, which outputs were generated, and what level of review occurred.
Appropriate use of AI inputs: Management should understand when AI is being used for exploratory support versus when its outputs are informing decision-sensitive work, board materials, partner discussions, regulatory activities, or quality-controlled records.
Practical enablement: The goal is to let scientists move faster while giving leadership confidence that AI use is governed, defensible, and aligned with the company’s risk posture.
We believe the right governance posture is a tiered, risk-based approach. Low-risk exploratory use should be easy. Sensitive use should have clearer guardrails. Regulated or decision-critical use should require additional review. This is how IT and cybersecurity can enable scientists without becoming blockers.
| Area | Management framing | Governance implication |
|---|---|---|
| Local project data | Conversations and artifacts are stored on the user’s computer rather than in a central Anthropic-hosted project store. | The endpoint becomes a primary control point. Managed devices, disk encryption, backup expectations, offboarding procedures, and clear storage rules are essential. |
| Model prompts and responses | Prompts and Claude responses are sent to Anthropic when the app calls the model. | Users need clear rules on what may be included in prompts. Sensitive IP, human subject data, clinical data, partner data, unpublished research, and board-sensitive information require explicit handling decisions. |
| Remote compute | Claude Science can send code and data to company-controlled servers, SSH hosts, HPC environments, or cloud accounts. | The destination environment needs appropriate identity, access control, storage governance, logging, cost management, and data classification. |
| Connectors | Directory, local, and custom connectors can give Claude Science access to scientific databases or other data sources. | Connector permissions, scopes, tunnels, and approved data sources need review before enablement. A connector can quickly turn an AI tool into an access pathway across sensitive systems. |
| Audit and compliance | Some audit, compliance export, and organization data export capabilities are not currently available for Claude Science data stored on member devices. | Do not assume standard SaaS audit, eDiscovery, or compliance exports cover Claude Science. Endpoint controls, local logging expectations, and clear policy become more important. |
| HIPAA and PHI | Anthropic’s admin documentation indicates that HIPAA-ready organizations can enable Claude Science beta, but Claude Science usage is not covered under the BAA. | PHI should be treated as out of scope unless the contract, workflow, and control environment are explicitly reviewed and approved. |
| Global privacy laws | Biotech organizations may be subject to GDPR, PIPEDA, PIPL, state privacy laws, clinical trial obligations, and contractual privacy commitments. | AI governance should address personal data broadly, not only PHI. Employee data, investigator information, participant data, genomic data, site data, vendor data, and cross-border processing all need to be considered. |
| Admin controls | Claude Science supports some enterprise controls, including SSO, SCIM, member management, roles, groups, model access controls, and usage analytics, while other controls are partial or still maturing. | Claude Science should be rolled out with a control model that reflects the maturity of the product. A beta product can be useful, but rollout should match available administrative and compliance capabilities. |
The reviewer and provenance model
Claude Science includes a reviewer that checks whether Claude’s claims match the approved plan, saved artifacts, and execution record. It can flag issues like unsupported citations, values that contradict source files, references whose DOI resolves to a different article, claims that work was run when it was not, or conclusions not supported by the method used.
This is valuable, but it is not a replacement for scientific review. Anthropic explicitly says the reviewer does not re-run analyses and does not decide whether the method was the right method for the research question.
Practical takeaway: provenance and reviewer checks make Claude Science slightly more “defensible” than ordinary chat output, but they do not eliminate the need for human scientific judgment, QA review, validation, or independent replication.
Use Case Landscape for Biotech
Claude Science is most compelling in areas where scientific work is already fragmented across tools, databases, compute environments, and documentation.
High-value use cases include:
Literature review and evidence synthesisClaude Science can query scientific sources, organize findings, and build reusable review workflows. This is especially useful for target assessment, disease-area landscaping, competitive intelligence, and indication dossiers.
Bioinformatics analysisIt can write and run Python, R, and shell code, create task-specific environments, and support domains like single-cell analysis, genomics, proteomics, and structural biology.
Scientific figure generationClaude Science can generate figures and preserve code, environment details, and execution history, making outputs easier to inspect and reproduce.
HPC workflow assistanceIt can help prepare and submit remote jobs to workstations or SLURM clusters over SSH. This could reduce friction for scientists who are less comfortable with terminal-driven compute workflows.
Reusable scientific workflowsTeams can turn repeated methods into skills, creating a shared library of research workflows rather than relying on one-off prompting.
Cross-functional scientific narrativesBecause it can combine analysis, artifacts, manuscripts, LaTeX previews, and provenance, it may help bridge raw analysis and boardroom-ready scientific communication.
Rollout questions biotech should be asking
Before enabling Claude Science broadly, companies should answer:
Who is allowed to use it?Computational biology and chemistry? Translational? Clinical? Business development? External consultants?
Have we mapped data flows and restricted boundaries?
What data is allowed?Public data only? Internal research data? Preclinical data? Partner data? Clinical data? Human subject data? PHI? GxP records?
Where may it run?Local laptops only? Managed Linux workstations? Approved HPC clusters? Modal? Other cloud VMs?
What compute identities are used?Individual SSH keys? Service accounts? Shared lab accounts? Federated access? How are permissions reviewed?
Which connectors are allowed?Public read-only connectors only? Directory connectors? Custom connectors? Internal APIs? ELNs? LIMS? Data lakes?
Who can create or install skills?Individual scientists? Research informatics? IT? A review board? Should skills be version-controlled?
What outputs are considered official?Are Claude Science artifacts exploratory only, or can they become part of scientific records, regulatory work, or quality-controlled deliverables?
How is local data governed?Are devices encrypted, backed up, monitored, and covered by offboarding procedures? Anthropic notes that removing a member does not wipe Claude Science data already on the user’s computer.
What is the policy for PHI and clinical data?Given Anthropic’s documentation that Claude Science beta is not covered under the BAA, PHI should be treated as out of scope unless explicitly addressed contractually.
How will scientific validity be reviewed?The reviewer can check claims against the execution record, but it does not replace method selection, scientific interpretation, QA review, or independent validation.
Recommended rollout model
For most biotech companies, the safest path is a staged rollout.
Phase 1: Controlled exploration
Start with a small group of scientists and computational users. Limit use to public data, non-sensitive internal test data, or approved synthetic datasets. Disable or restrict risky workflows through policy and device controls where possible. Define what may and may not be pasted into prompts.
Phase 2: Governed scientific workflows
Identify two or three repeatable high-value workflows, such as literature review, target assessment, single-cell exploratory analysis, or figure generation. Turn these into documented, reusable skills. Establish review expectations for outputs.
Phase 3: Approved compute integration
Connect only approved compute environments: managed workstations, approved HPC login nodes, or controlled cloud compute. Review SSH access, filesystem permissions, job submission behavior, output storage, and cost controls before enabling broader use.
Phase 4: Internal connector strategy
Evaluate whether Claude Science should connect to internal tools such as ELN, LIMS, data lakes, knowledge bases, or proprietary pipelines. Treat custom connectors as software integrations requiring security review, access design, logging, and data classification.
Phase 5: Enterprise governance
Move from experimentation to operating model. Define ownership across Research, IT, Security, Legal, Privacy, and QA. Create policy, training, approved use cases, restricted use cases, review checkpoints, and offboarding procedures.
Bottom line
Claude Science is not just another AI chatbot for scientists. It is an early version of an AI-native scientific workbench: local-first, compute-aware, skill-driven, connector-enabled, and built around reproducible artifacts.
For biotech companies, the upside is substantial: faster analysis, better literature synthesis, more reusable workflows, easier access to scientific databases, and lower friction between scientific questions and computational execution.
The risk is equally real: local data sprawl, immature audit/compliance controls, prompt-retention considerations, custom connector governance, remote compute exposure, cost management, and ambiguity around when AI-generated scientific work becomes part of an official record.
The right approach is not to block it reflexively or enable it casually. The right approach is to treat Claude Science as a new scientific computing surface — one that deserves the same thoughtful governance as cloud research environments, ELNs, LIMS, source control, and high-performance computing.
