LLM and MCP: wiring a model into the SI without losing control
RAG to read docs is one job. Acting on tickets or stock is another. MCP is a bounded way to expose tools to a model.
A RAG chatbot is enough when you only need to read documentation. As soon as the model must act — open a ticket, read stock, pull a contract — people go back to copy-paste.
What MCP actually is
Model Context Protocol: a way to let an LLM call tools (ticketing, ERP, DMS, logs) with bounded rights. The model does not wander through the IT landscape. It calls what developers exposed.
Why developers care
They do not glue a plugin per AI vendor. They expose tools once, with scopes and a log. That is SI integration, not decorative prompting.
Data
It does not have to go through a public LLM. You choose the model, the region, what leaves. MCP tools stay with you (or in the EU). That is scoping: hosting, journals, who may call what.