An MCP integration still needs the same approval gates a custom API integration needs

Model Context Protocol solves a real, narrow problem: it gives an AI agent one consistent way to call into Workday, Salesforce, SAP, or any other system, instead of a bespoke integration per system per agent. That's a genuine engineering win — it's not a governance win by itself, and pitches that imply MCP alone “replaces integrations” or removes the need for controls are overstating what the protocol does.

Our own MCP demo enforces a hard per-IP rate limit (20 requests) on the endpoint before a single call reaches the model — a basic control an MCP connection doesn't give you automatically. The system prompt also requires the model to return which specific tool it called and with what data, in a structured format the application parses back out — so “the agent answered a question” is always paired with “here is exactly which system record it touched to answer it.”

Neither of those is an MCP feature. They're governance decisions that have to be made explicitly, the same way they would for a hand-rolled REST integration — MCP standardizes the wire format, not the authorization model on top of it.

Not “do you support MCP” — almost everyone will soon. Ask what happens when the agent asks for something out of its granted scope, what the rate and spend ceilings are, and what evidence exists after the fact showing which tool call produced which answer. If the answer is “MCP handles that,” it doesn't — the protocol carries the request, not the policy.