NetBox MCP Server - Query the Source of Truth
Solution overview
An AI agent that doesn't know your network has three options when you ask about it: refuse, hedge, or make something up. None of those answers can feed an automation decision, because none of them traces back to a system of record.
In this lab, you catch that problem yourself and then fix it. You start by asking an AI agent about a Cisco Nexus VXLAN/EVPN fabric with no source of truth connected, and you capture the answer as a baseline. Then you register the pre-installed NetBox MCP server with the agent. The Model Context Protocol (MCP) lets the agent query NetBox as a set of tools instead of answering from memory. Re-ask the same question and compare the two answers.
From there, you build up from single lookups to chained, multi-tool queries that answer compound operational questions about the fabric's spines, leaves, servers, and VXLAN overlay networks. You finish by producing a structured, schema-conformant intended-state report. You then spot-check that report directly against NetBox.
Every claim the agent makes maps to a query you can re-run, and that traceability is what makes its answers usable. The intended-state report you build is the same kind of input a drift detector or remediation workflow can trust.