Design an MCP server that gives an AI assistant access to a customer's internal ticketing system and knowledge base, for an enterprise with strict access controls.
Answer it out loud first, then compare. That's what makes it stick.
I'd clarify who the users are and what actions are needed, probably read tickets, search articles, and maybe add comments. I'd expose few, workflow-shaped tools, read tools open and write tools gated. Auth uses the user's own identity through the company's identity provider, so the assistant never sees more than the user could. Every call is logged, results are trimmed, and writes need approval. I'd measure it on real questions graded by their staff.
- Clarify users, actions, sensitivity, identity provider
- Small tool set: search, get, add_comment; no bulk export
- Auth: per-user tokens scoped to this server, no passthrough
- Read tools open, write tools behind approval
- Audit log of every call with user and arguments
- Eval on real questions; staged rollout
Clarify first (2 minutes). Who uses it (support agents, engineers, everyone)? Which actions: read, search, comment, close? Which tickets are restricted (HR, security, legal)? Which identity provider, and may ticket text leave their cloud?
Tool surface
tickets_search(query, filters, limit)returning short summaries with ids, not full bodies.tickets_get(id)returning the fields needed to reason, with long threads truncated and a note on how to page.kb_search(query)andkb_get(id).tickets_add_comment(id, text)as the only write, marked non-read-only. Notice what is absent: no delete, no bulk export, no free-form query language.
Identity and authorization
- Use OAuth with the user's identity so every call is made as that user and the ticketing system's own permissions apply. The assistant should never get a broader service account that bypasses row-level rules.
- The server must accept only tokens issued for it (audience check) and must not pass the client's token on to the ticketing API; it obtains a separate downstream credential.
- Filter by permission at query time, not after retrieval, so restricted ticket text never reaches the model.
Safety
- Treat ticket text as untrusted input: customers can type instructions into tickets.
- Write tools need approval in the host; log the exact comment text shown to the approver.
- Rate limit per user and per tool.
Operations
- Structured audit log: user, tool, arguments, result size, latency.
- Alerts on denied calls and unusual volume.
- Return actionable errors ('result truncated: add a status filter').
Evaluation and rollout. Build 100-200 real questions with answers staff can verify, include permission-leak tests (a user asking about a ticket they cannot see), and roll out read-only first, then add comments for a pilot team.
- A user asks about a ticket they are not allowed to see. Walk through what happens.
- How would you handle a ticket with 3,000 comments?
- How do you test the permission model before launch?
- Forward Deployed Engineer Interview Prep (2026) - DataInterview (prep, 2026-02-13) Includes the prompt 'Design an MCP server that gives Claude access to a customer's internal ticketing system and knowledge base' as a design practice question.
- MCP Agent Interview: Internal Tools Case - DataInterview (prep, 2026-10-12) Describes an Anthropic-style use-case round with MCP over internal tools, judged on reliability and context management, with per-server tokens and host-side approvals.
- Authorization - Model Context Protocol specification (docs, 2025-06-18) Servers must validate token audience and must not pass received tokens through to upstream APIs.