MCP Server Permissions Checklist for PHI: SharePoint, Google Drive, Postgres and Snowflake
Give each MCP server its own identity, grant that identity only the PHI the task needs in each source platform, restrict which tools and operations the agent can invoke at runtime, and record every tool call as evidence. Source-system permissions set the ceiling on what an agent can reach. They do not decide whether a specific tool call should be allowed, redacted, blocked or escalated.
- Each MCP server has its own identity. No shared service accounts across servers, environments or platforms.
- Where the agent acts for a user, access is delegated in that user's context so the agent cannot exceed the user's own permissions. Where a service identity is unavoidable, it is scoped to a defined task.
- Default access is read-only. Write, delete, share and export rights are granted only with a documented reason.
- Credentials are short-lived where the platform allows, held in a secrets manager and rotated. They never appear in prompts, tool arguments or configuration files.
- OAuth grants and API scopes are the narrowest set that supports the tool's function, and broad or tenant-wide grants are reviewed and removed.
- Each server has a named owner, a documented data scope and a recorded decision on its business associate status.
- Tools are on an allow list. Any tool or operation not listed is denied by default.
- Operations with high impact, such as bulk export, sharing changes, deletes and writes, are blocked or limited to named tools.
- Responses are inspected, and PHI the task does not need is redacted before it reaches the model.
- Rate limits apply per identity and per tool, to constrain bulk retrieval and enumeration.
- Defined high-risk calls escalate to a human approver, and the approval, approver and outcome are recorded.
- Policy decisions of allow, redact, block or escalate are tested against known cases before the server goes live and again after any change.
Why source-system permissions are necessary but not sufficient
The HIPAA Security Rule establishes national standards to protect electronic protected health information that a covered entity or business associate creates, receives, uses or maintains. Technical safeguards sit at 45 CFR 164.312, which, as of November 2026, is published in the eCFR as current law. Audit controls are addressed at 164.312(b). An MCP server that reads from a PHI-bearing repository is part of that environment, so its access has to be scoped and its activity has to be evidenced.
The common failure is inheritance. An MCP server is often wired to a shared service account or a broad OAuth grant, and the agent behind it can then reach whatever that identity can reach. The source platform will honor those grants correctly. It has no view of whether a particular tool call, made for a particular task, needs the data it returns.
This checklist therefore keeps three layers apart. The first is what is configured in the source system. The second is what policy is enforced at runtime on each tool call. The third is the evidence that shows both. A control that exists only in the first layer should not be recorded as satisfying a requirement that belongs to the second. Also confirm, under your own agreements, whether any party hosting or operating an MCP server is a business associate.
Data-level access by platform
Data-level controls decide which records and fields an identity can read. Configure them in each platform, then test them from the MCP server's identity, not from an administrator account. A permission review done as an administrator will miss inherited access.
The four platforms use different models, so the items below are specific to each. In every case the goal is the same: the server's identity can see only the repositories, tables and rows the task requires, and nothing is reachable by default through group membership, sharing or ownership.
- SharePoint: Grant access at the specific site or library level, not tenant-wide. Check for inherited permissions, broad organizational groups and anonymous or organization-wide sharing links on PHI libraries. Exclude libraries the task does not need.
- Google Drive: Request the narrowest OAuth scopes the tool needs. Limit the identity to specific shared drives or folders. Review link sharing on PHI content, and avoid domain-wide delegation unless a documented need exists.
- Postgres: Use a dedicated role with no superuser rights and no object ownership. Grant SELECT only on named schemas, tables or views. Apply row-level security policies to that role, and expose views that omit direct identifiers the task does not need.
- Snowflake: Use a dedicated role granted only the required databases, schemas and views. Apply masking policies to PHI columns and row access policies to restrict records. Assign a separate warehouse with resource limits.
Evidence and retention for audit controls
Audit controls at 164.312(b) concern recording and examining activity in systems that contain or use ePHI. For an MCP deployment, the practical question a reviewer will ask is what the AI was allowed to do and what it did. Source-platform logs show that a service account read data. They rarely show which agent, which tool call, under which policy, with what result.
Define the evidence record before launch. At minimum, capture the calling identity and agent, the tool and its parameters, the policy decision applied, any redaction performed, any approver and approval outcome, the timestamp and the result. Keep this separate from the source system's own logs, protect it against alteration, and set retention according to your own policy and legal advice.
Then test the evidence. Pick a sample of tool calls and reconstruct the story from the record alone: who initiated it, what was permitted, what was returned, and whether anything was redacted or escalated. If the record cannot answer those questions, the control is incomplete.
Where Trussed fits
Trussed enforces policy in the execution path for MCP servers and tool calls, with allow, redact, block, rate limit and escalate-for-approval decisions applied at runtime, not reviewed after the fact. It generates audit-ready evidence mapped to HIPAA, and it is deployed self-managed in the customer cloud with configurable retention and data residency. Trussed runs in production with healthcare organizations. It complements source-system permissions and does not replace them.
Three layers to verify for every MCP server touching PHI
Source-system permissions
Identity, grants and data-level controls configured in SharePoint, Google Drive, Postgres or Snowflake.
Runtime policy
Allow, redact, block or escalate decisions made on each tool call in the execution path.
Evidence record
A durable record of what the agent was allowed to do and what it did.
Check your MCP servers against the runtime layer
See how policy is enforced on tool calls and how audit-ready evidence is produced for PHI access.
See How Evidence Is Produced