How would you roll out MCP (Model Context Protocol) servers across an enterprise securely?
Short answer
Treat MCP servers as APIs: maintain an approved registry, run them behind a gateway with authentication (OAuth with per-user delegated access), least-privilege tool scopes, review of tool descriptions, logging and rate limits, separate read and write tools with approvals for writes, and version and test them like any production service.
Detailed answer
- Registry: approved MCP servers with owners, data classification, risk tier.
- Identity: users authenticate; MCP servers call backends with delegated tokens scoped to the user (no shared super-credentials).
- Gateway: central policy enforcement, allowlists per AI application, rate limits, audit logs.
- Tool hygiene: concise, accurate tool descriptions (to prevent tool poisoning or misuse), input validation, output size limits.
- Writes: confirmation steps; human approval for high-impact actions.
- Lifecycle: versioning, testing with agent eval suites, deprecation.
What the interviewer expects
Applying API security and governance principles to agent tooling.
Common mistakes to avoid
Allowing any developer-built MCP server to connect to production data with broad credentials.