Model Context Protocol (MCP)

GLOSSARY

Definition

The Model Context Protocol is an open standard for connecting AI models to external tools and data sources. An MCP server exposes capabilities - search a database, send an email, read a file - in a uniform format any MCP-compatible client can use. It replaces one-off integrations between each model and each tool with a single protocol on each side.

Why it matters

Anthropic introduced MCP in late 2024 and open-sourced it; adoption spread quickly through developer tools and then marketing platforms. For marketing teams, MCP is how agents get hands: the CRM connector, the ad-account connector, the analytics connector that lets an agent act inside your stack.

How it works

An MCP server is a small program that declares resources it can read and actions it can perform. The client - Claude Code, an agent runtime, an IDE - discovers those declarations and lets the model call them as structured tools. Sessions are stateful enough for context to persist across calls within a conversation.

Practical uses

Marketing teams use MCP servers to give agents safe access to internal data: campaign performance reads, content draft writes, ticket lookups. The protocol's permission model means access can be scoped per server, which is how you give an agent analytics reads without handing over spend controls.

How to choose

Prefer MCP servers that are idempotent and read-heavy for first deployments. A server that can only read and report cannot break production data; once trust is established, add write-capable servers one at a time.

The numbers

Scale math: an agent that checks six data sources before each decision, running once per hour, makes roughly 4,300 tool calls per week per agent. Against metered MCP or search quotas, that is the difference between a rounding error and a budget line - design polling frequency before launch, not after.

Common mistakes

Running unvetted third-party MCP servers with production credentials is the emerging horror story - a server with broad scopes is a supply-chain risk. Audit what scopes each server holds and rotate credentials separately.

What changed with AI

MCP exists because of AI; the practical risk is quota economics. Hosted-model providers meter MCP tool calls separately from plain inference, and agentic workloads multiply call counts. Know your provider's metering before wiring an agent to a chatty tool.