AI-Assisted Operations Platform
The problem
MSP work is swivel-chair work. A single client question routinely touches the ticketing system, the documentation platform, the RMM, the network controller, the firewall manager, and a licensing portal. Each has its own portal, its own auth, its own query language. The information exists; the hours go to collecting it.
The approach
The Model Context Protocol gives an AI agent a standard way to call tools. So instead of an engineer visiting six portals, I built a custom MCP server that exposes the vendor APIs an MSP actually runs on as tools Claude can call: ticketing, CRM, documentation, RMM, network management, firewall management, licensing, and monitoring.
One conversational request like "what changed on this client's network this week, and are there any related tickets" becomes parallel API calls, cross-referenced, with the answer synthesized and every claim traceable to a source system.
Build notes
- Single Node.js MCP server with a module per vendor, all following one reference pattern: auth fallback chain, automatic retry on expired sessions, consistent result wrapping
- Each vendor module declares its tools with JSON Schema so the agent knows exactly what parameters are legal before it calls
- Read operations are the default; anything that writes to a production system carries explicit guardrails
- An evidence rule sits on top of everything: every statement in a generated report must trace to an API response, not model inference
Outcome
Cross-platform questions that took an afternoon of portal-hopping now resolve in one conversation. Client health checks, network change reviews, license audits, and incident timelines pull from live systems instead of stale exports. The platform also became the foundation for downstream automation: once the tools exist, every new workflow starts closer to done.
Lessons
- The reference pattern matters more than any single integration. Module ten took an hour because modules one through nine established the shape.
- AI agents need guardrails at the tool layer, not just the prompt layer. A prompt can be talked out of a rule; a tool that refuses to write to production cannot.
- Evidence discipline is what separates useful AI output from confident fiction. Enforce it structurally.