· Gary O'Keeffe · AI & Data · 6 min read
How to Connect Claude Enterprise to Your Internal Lab Data
Claude runs in Anthropic's cloud. Your LIMS and file shares don't. Here are the secure options that exist today.
Your team has Claude Enterprise. Within a week, someone asks the obvious question: “Can it look at our assay results?” Then IT asks the next one: “How, without punching a hole in the firewall?”
This guide covers the options that exist today, what each one can and can’t do, and the pattern we recommend for a growing life science company.
Where Claude actually runs
Claude Enterprise is SaaS. When Claude uses a connector, the request comes from Anthropic’s cloud, not from the user’s laptop. Your server must be reachable over the public internet from Anthropic’s infrastructure. Servers on a private corporate network, behind a VPN, or blocked by a firewall won’t connect directly—though MCP tunnels are available by request for Enterprise organizations that need to reach private networks.
Your LIMS, ELN, and instrument file shares are private by design. So the question isn’t “how do I plug Claude in?” It’s this: what should Claude be allowed to see, and what’s the safest path to it?
Option 1: Native connectors
Claude has Anthropic-hosted connectors for Microsoft 365 (SharePoint, OneDrive, Outlook, Teams) and Google Workspace (Gmail, Calendar, Drive). On Enterprise, an Owner enables them for the organization, and each user then connects their own account. See the connectors directory for the current list.
The connectors use delegated permissions, so users see only what they can already see. Review sharing settings before you enable this, and leave write tools off until you have a specific reason to enable them.
Native connectors are a good first step for documents. They don’t reach your LIMS, instrument data, or network drives.
Option 2: A custom remote MCP connector
The Model Context Protocol (MCP) is the open standard Claude uses to call external tools. Enterprise Owners can add a custom connector by pointing Claude at an MCP server URL. Users then connect individually.
This is how you reach lab data. It also raises three security questions you must answer properly.
Where does the server live?
Don’t put an MCP server on the lab network and open a port. Run it in your own cloud account (AWS, Azure, or GCP), next to curated copies of your data, and expose only the MCP and OAuth endpoints through a managed gateway or WAF.
On AWS, you can use API Gateway with Lambda or a container-based MCP server. On Azure, API Management can front your MCP server and enforce OAuth/JWT validation. Choose the gateway that fits your existing cloud infrastructure.
Who is the user?
Use per-user OAuth through your identity provider (Entra ID, Okta, Google). Claude also supports Enterprise Managed Auth for SSO-based connection without a per-user consent screen. With a shared API key, everyone who uses the connector reaches the service with that same credential—that’s generally not acceptable for sensitive data.
One detail to note: Anthropic’s egress range is shared across its services. An IP allowlist admits Anthropic, not your specific tenant. It’s a useful filter, not authentication.
What can the tools do?
Keep tools read-only and narrow: search_samples, get_slide_metadata, or query_assay_results against approved views. Enforce per-user permissions inside the server. For regulated environments, avoid giving Claude tools that write to systems of record.
Option 3: Local MCP on Claude Desktop
Claude Desktop can run local MCP servers on the user’s machine. A local server can reach whatever that laptop can reach, which is exactly why it’s harder to govern at scale. Consider reserving this for a few power users with read-only credentials, and review your organization’s policies on local tool execution.
Option 4: Run Claude inside your own cloud
If data must stay in your cloud account or a specific region, use Claude through Amazon Bedrock or Google Vertex AI. Bedrock supports PrivateLink for private VPC connectivity. Vertex AI offers regional endpoints and VPC Service Controls.
The trade-off: you get the model, not the Claude Enterprise app. You’ll need a simple internal interface and your own retrieval layer. For many growing biotechs, this is the answer for a specific restricted dataset, not for general use.
Don’t skip the boring part
Residency. For usage-based Enterprise plans, Anthropic offers US-only inference, which keeps model processing within the United States. This setting applies to inference, not to data storage or third-party connector processing. For data that must stay in specific regions, consider Option 4.
Audit. Enable the Compliance API early. Its Activity Feed retains events for six years but isn’t retroactive—the clock starts when you enable it. Also log every MCP tool call at your gateway with the user’s identity, and send both to your SIEM.
A note on regulated environments
Treat Claude as decision support reading controlled copies, not as a system of record. If Claude’s output informs a regulated decision, document the intended use, assess the risk, and keep a human in the loop. Involve QA early. We don’t provide regulatory advice, but we can help you design systems with auditability and control in mind.
The pattern we recommend
Fix the data first. Inventory it, link raw files to LIMS/ELN records, and land a curated, read-only layer in your cloud. For a digital pathology company, we did exactly this for about 1 PB of slide images: we parsed headers and paths, linked them to the LIMS, and indexed them in OpenSearch. It runs for under $1,000 a year. That index is what Claude should query, not the raw share.
Turn on native connectors for a pilot group, after a sharing review, with write tools off.
Add one governed custom MCP connector over one curated domain: in your cloud, behind a gateway, with per-user OAuth, read-only tools, and every call logged.
Use Bedrock or Vertex AI only for data that can’t leave your cloud or region.
Revisit quarterly. These products change monthly.
Checklist
- SSO enabled for Claude Enterprise
- Compliance API enabled; events flowing to the SIEM
- US-only inference decided (if eligible and required)
- SharePoint/Drive over-sharing reviewed before connectors are enabled
- Write tools off by default
- MCP server in your cloud, not on the lab network
- Per-user OAuth via your IdP; no shared API keys for sensitive data
- Read-only tools over curated views
- Gateway logs every tool call with user identity
- GxP position documented with QA
Want this designed and running in weeks? Our AI Data Readiness service starts with a 1–2 week assessment and can put your first secure connector live in a 4–6 week pilot.
