How it works, minus the jargon
An MCP server is the product's side of the plug: a list of tools, each with a name, a description, and a typed schema for its inputs. An MCP client is the agent's side — Claude, an IDE agent, a custom harness. When they connect, the agent reads the tool list and can call any of them: fetch a report, create a segment, draft a variant. Authentication is typically a scoped token, so the agent holds exactly the permissions it was granted and nothing else.
The point of the standard is that neither side needs to know about the other in advance. Any MCP-capable agent can drive any MCP server, the way any browser can load any website.
Why marketers should care
MCP quietly turned "can my agent drive it?" into a procurement question. If your team works alongside agents, a platform with no MCP surface is a platform your agents can't touch — every automated workflow dead-ends at a human clicking through a UI. Tool coverage now matters the way "has an API" mattered a decade ago, and the delta between "has an MCP server" and "an agent can genuinely run the product" is large. Ask for the tool list.
AXO's MCP server exposes 130+ tools spanning the full personalization loop — site crawling, segments, variants, placements, triggers, lead capture, pipeline, and holdout measurement — hosted, with scoped tokens, and mirrored over REST.