Integration
Free Claude Certified Architect - Professional practice — 6 questions on Integration, with explanations. No sign-up. Full 12-question mixed test →
A healthcare company wants Claude Desktop users to query patient records through an MCP server connected to their EHR system. Different clinicians have different access levels based on role (nurse vs. physician vs. billing staff), and every access to PHI must be individually auditable. Which authentication/authorization pattern should the architecture use?
Token passthrough with server-side RBAC enforcement preserves per-user identity, enables fine-grained audit logging tied to the actual clinician, and keeps authorization as a server-side control rather than a client-side or model-side assumption.
A Claude-powered agent processes incoming webhook events from a payment processor to update order status in an ERP system via an MCP tool. The payment processor's webhook delivery is at-least-once, so duplicate events are expected during network retries. How should the integration be designed to prevent duplicate ERP updates?
Idempotency keys based on a unique event ID, checked against a persisted store before invoking the downstream tool, is the standard pattern for safely handling at-least-once delivery semantics regardless of Claude's involvement.
An enterprise's core order-management system only exposes a SOAP/XML interface over a mainframe, and security policy prohibits opening any new network path directly to the mainframe for AI tooling. The architect needs Claude to read and update order records via MCP. What is the best integration approach?
Introducing a middleware/adapter layer respects the existing security policy (no new direct network path to the mainframe), provides a clean abstraction boundary for the MCP server, and allows modern auth, rate limiting, and monitoring to be layered in front of the legacy system — the standard enterprise pattern for legacy constraints.
A retrieval pipeline must support both exact lookups of high-cardinality part numbers (millions of unique SKUs) and semantic search over free-text part descriptions, surfaced to Claude via a retrieval tool. Which retrieval architecture best satisfies both requirements?
Hybrid retrieval combining exact-match structured lookups for high-cardinality identifiers with vector search for unstructured text is the correct enterprise pattern — vector similarity is unreliable for exact-match requirements like SKUs, while exact-match indexes can't handle semantic description queries.
A nightly batch job needs to process 500,000 documents through the Claude API for classification, writing results into downstream MCP-connected systems, without overwhelming rate limits or downstream services. Which design best handles this at enterprise scale?
A queue-based worker pool with adaptive concurrency driven by rate-limit headers and exponential backoff on throttling responses is the standard backpressure-aware pattern for large-scale batch integration, maximizing throughput while respecting both the Claude API and downstream system limits.
An MCP server exposes both read (query) and destructive (delete) tools against a CRM system. The architect wants Claude to answer sales questions conversationally but must prevent any possibility of accidental data deletion triggered through conversation, even if a user asks Claude to delete something in a confusing or adversarial way. What is the best architectural control?
The least-privilege principle requires that destructive capabilities simply not be reachable from a context where they shouldn't be used, rather than relying on prompts, conversational confirmations, or after-the-fact remediation — all of which are model-dependent or reactive controls, not enforceable security boundaries.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →