The problem
Agents needed to research across customer history, transactions, balances, account status, prior tickets, policy, and internal documentation before they could confidently answer many customer questions.
What I built
Using Claude, I built a custom MCP server that called Zendesk directly through its API. I added integrations with our internal databases so the workflow could use a participant code and ticket ID to assemble a fuller customer picture: transaction history, balances, account status, ticket history, and other relevant context.
Zendesk Explore did not provide open API endpoints for everything I wanted, so I reconstructed useful reporting from the data points that were available. Where that did not make sense, I kept using exports rather than pretending every limitation needed an elaborate workaround.
Human review was intentional
The system could generate a recommended response based on the research it had done plus extensive markdown guidance I created for tone, language, compliance, sensitive issues, and when to be friendly versus formal. I deliberately did not give it permission to send responses. The agent saw the research, reviewed the draft, changed anything they wanted, and sent it themselves.
What this does and does not prove
I am not a traditional software engineer. I cannot sit down and code this from memory without AI. I can design the system, read API documentation, work comfortably in a CLI, reason through data and permissions, use coding agents to implement it, and keep debugging until it works. That distinction matters to me.