MCP servers in production: what our internal platform taught me
We built an MCP service layer over the APIs we already trusted, and let support and engineering ask questions in plain language. Response times fell 40 percent. The lessons were about interfaces and guardrails, not models.
The most useful AI system I have shipped is not a chatbot. It is a service layer. We built an internal support platform on MCP, the Model Context Protocol, so that support agents and engineers can query logs, account state, and knowledge bases in real time through tools the model can call. Response times fell 40 percent and engineering escalations fell about 45 percent. The interesting part is how little of that outcome came from the model.
1. Start from the APIs you already trust#
We did not build new backends for the AI. The MCP layer sits in front of the same internal APIs that our dashboards and support tooling already used, with the same authentication and the same audit trails. Each tool is a thin, named operation over an existing capability: fetch this account's recent events, search the knowledge base, pull these logs for this window.
That decision did two things. It kept the security review honest, because no new data paths appeared. And it meant the tools behaved predictably, because they were already battle-tested by humans using them through other screens.
2. Tools are an interface, design them like one#
A model calling your tool sees its name, its description, and its parameters, and nothing else. Vague descriptions produce vague calls. We learned to write tool descriptions the way you would write API documentation for a smart but literal colleague: what it returns, what the units are, what an empty result means, and when not to use it.
Small, composable tools beat a few mega-tools. A search tool, a fetch tool, and a summarize-my-input instruction from the caller will outperform one giant do-everything endpoint, because the model can chain small certainties instead of betting on one big guess.
3. Guardrails matter more than the model#
Read-only by default is the rule that saved us the most future pain. The platform can look at almost anything the caller is allowed to see, and change nothing. Every call is authenticated as the human asking, authorized against the same permissions, and logged with the question and the tools used.
Data boundaries need the same care as any API. Tools return the minimum fields the task needs. Customer data that support should never see does not become visible just because a model asked nicely.
4. What changed for the team#
The win was not that a model answers questions. It is that the first look at a problem, gathering the logs, the account history, and the relevant runbook, now takes seconds instead of a tab-switching expedition. Escalations fell because support could resolve the routine cases with the same facts engineering would have gathered.
If you are starting an MCP project, my advice is boring on purpose: wrap the APIs you already trust, write descriptions like documentation, keep it read-only until you can prove you need more, and log everything. The model is the easy part to swap later. The interface is the product.
