Book a demo
Topics · MCP

Insurance MCP Governance Access Control | Moring

Access control for MCP governance in insurance decides which agents may reach which MCP servers and tools, under whose credentials, and with what limits before a call executes. Because insurance agents touch policy administration systems, claims records, and rating engines holding nonpublic personal information, access control has to bind each agent to its own scoped identity rather than a developer's token.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Access control for MCP governance in insurance decides which agents may reach which MCP servers and tools, under whose credentials, and with what limits before a call executes. Because insurance agents touch policy administration systems, claims records, and rating engines holding nonpublic personal information, access control has to bind each agent to its own scoped identity rather than a developer's token.

How it works

Access control inside MCP governance is the discipline of deciding, before any agent call executes, which Model Context Protocol servers an agent may reach, which tools it may invoke, and under whose authority the call runs. In an insurance environment the stakes are specific: MCP servers connect agents to the policy administration system, the claims platform, the rating engine, and the customer records behind them, exactly where nonpublic personal information and, for health lines, protected health information live. An ungoverned connection at that boundary is not a productivity detail; it is an authority decision about who can read and write the systems that move the combined ratio.

The structural problem is that insurance cores were not designed for agent traffic. The dominant policy administration cores, including Guidewire and Duck Creek, and a long tail of legacy systems were not built for fine-grained delegated access, so an agent that connects through an MCP server often inherits whatever credential the integrating developer had. Research on MCP deployments reports that 53 percent of MCP servers rely on static API keys, which means the agent acts with a human-shaped identity that cannot be scoped, rotated, or attributed to the agent itself.

Access control for insurance therefore starts with per-agent identity. Each agent receives credentials issued to the agent, not borrowed from a person, with scope limited to the servers and tools its workflow requires. When an MCP server supports authorization over HTTP, the specification calls for OAuth 2.1 with PKCE, protected resource metadata, and audience-bound tokens, and a governance layer should enforce that standard on its own MCP endpoint. A claims status agent then holds read scope on claim records while a payment instruction tool stays outside its reach.

Scoping has to operate at the tool level, not the server level. Governance guidance for MCP is explicit that a principal who may reach a server should not automatically be able to call every tool on it, because servers expose capabilities of very different weight, from document search to payment initiation. For an insurer this distinction maps cleanly onto claims work: an agent summarizing a file may read, while an agent issuing a settlement instruction needs a separate, narrower grant.

Enforcement belongs at a single runtime point. A gateway in front of MCP servers evaluates each call against policy before execution, checking the agent identity, the requested tool, the arguments, and the risk signals, and returns an allow, deny, redact, or human-approval decision. Insurance adopters of that pattern map the controls directly to their regulatory obligations, including NAIC expectations for insurer AI programs, so the same decision point produces both the enforcement and the evidence.

Access control also has to survive change. MCP tools are dynamic: a server can add capabilities between one session and the next, which means a permission granted in March may cover a tool that did not exist then. Periodic revalidation of agent scopes against the current tool inventory keeps the grant aligned with what the server actually exposes.

Finally, access control is the control that examiners and courts will ask about first. When a governance lapse at a carrier surfaces years later in an exam or in litigation, the first questions are who was allowed to act, under which credential, and who decided that. Scoped per-agent access with recorded decisions answers those questions by construction rather than by reconstruction.

Frequently asked questions

How does access control work for MCP servers in an insurance environment?

Access control in insurance MCP governance intercepts each agent tool call at a gateway and checks the agent identity, the requested tool, and the arguments against policy before execution. Claims, policy administration, and rating systems stay behind that decision point, so an agent receives scoped credentials of its own rather than a developer's static key, and every allow, deny, or redact decision is recorded for review.

Why can an insurer not rely on shared API keys for agent access to core systems?

Static, shared credentials cannot be attributed to a specific agent, rotated on demand, or scoped to the tools a workflow actually needs, and 53 percent of MCP servers still use them. In an insurance setting that means an agent reaching policyholder records through MCP acts with an identity nobody can tie back to a governed process, which undermines both least privilege and the audit trail an examiner will request.

What does least privilege look like for a claims processing agent?

A claims agent in an insurance MCP deployment should hold read scope on the claim records it needs, write scope only on the fields its workflow updates, and no grant at all on payment initiation or rating engine configuration. Tool-level permissions matter more than server-level ones, because the same MCP server can expose both a document search tool and a settlement instruction tool with very different consequences.

How does access control support NAIC and EU AI Act expectations for insurers?

Regulatory scrutiny of insurer AI concentrates on accountability: who authorized the system, what it may do, and what evidence exists of its decisions. Per-agent identity, scoped tool permissions, and a recorded decision at every MCP call produce that evidence directly, which lets an insurance compliance team demonstrate governed access to policyholder data and consequential actions rather than describing controls that exist only as policy documents.

Sources

  1. tyk.io
  2. portkey.ai
  3. mcpmanager.ai
  4. arxiv.org
  5. ascendingdc
  6. getmaxim.ai
  7. kastra.ai
  8. - content.naic.org
  9. - artificialintelligenceact.eu
  10. - eur-lex.europa.eu