Every platform team has the same artifact: a wiki of correct procedures nobody reads. Documentation assumes a human will translate intent into a precise sequence of commands against the right cluster with the right flags. Most won't. They file a ticket, which costs the platform team a person, or they guess, which costs everyone an incident. The knowledge isn't missing. The interface is.
An MCP plugin platform and a repository of platform skills covering cluster and node operations — turning written procedure into executable, governed action. An engineer states intent; the skill carries the correct sequence, the guardrails, and the audit trail. The platform team writes the skill once instead of answering the ticket forty times.
The governance layer. It's straightforward to expose a tool to a model and much harder to keep authorization, blast-radius limits, and auditability intact when the caller is non-deterministic. A skill that can do something dangerous needs to be as constrained as the runbook implied but never enforced.
MCP quietly turns a docs site into an API for intent, and that reframes the whole internal-developer-platform category. The winning products won't be the ones with the most integrations — they'll be the ones where exposing a capability to an agent doesn't mean rebuilding your permission model from scratch.