Anthropic open-sourced the Model Context Protocol on November 25, 2024, and almost nobody in marketing noticed. Twenty months later, MCP sits underneath tools you may already be using, the ad industry’s newest transaction standard is built on top of it, and every major AI lab has adopted it. If you run a marketing stack, this protocol is now your plumbing whether you chose it or not.
The short version of what MCP is: a standard way for an AI model to connect to external tools and data. One connector spec instead of a custom integration for every tool-and-model pair. Before MCP, wiring an AI assistant into your CRM, your analytics and your ad platforms meant bespoke code for each combination. With MCP, a tool exposes one server, and any compliant AI client can use it.
USB-C is the tired analogy, but analogies get tired by being right.
From side release to industry standard in thirteen months
The adoption timeline is worth laying out, because I cannot think of another standard the big AI rivals agreed on this fast.
- November 2024: Anthropic releases MCP as an open standard, with SDKs and pre-built servers for Google Drive, Slack, GitHub and Postgres.
- March 2025: OpenAI adopts its main rival’s protocol across the Agents SDK, ChatGPT desktop and the Responses API. Sam Altman’s summary: “People love MCP.”
- April 2025: Google DeepMind commits Gemini models and its SDK to MCP.
- May 2025: Microsoft and GitHub join the MCP steering committee at Build, with support spanning GitHub, Copilot Studio, Azure and Windows 11.
- December 2025: Anthropic donates MCP to the Linux Foundation’s new Agentic AI Foundation, with AWS, Bloomberg, Cloudflare, Google, OpenAI and Block among the founding members.
When Google signed on, DeepMind’s CEO put it plainly:
“MCP is a good protocol and it’s rapidly becoming an open standard for the AI agentic era,” wrote Google DeepMind CEO Demis Hassabis in April 2025, reported by TechCrunch.
Competitors adopting a rival’s standard and then handing it to a neutral foundation is the arc HTML and Kubernetes followed. That arc ends with the standard being invisible and everywhere.
Why a media buyer should care
Two developments made MCP concrete for marketers rather than a developer curiosity.
First, the platforms started shipping servers. Google Analytics released an experimental open-source MCP server in July 2025 that lets an AI assistant query a GA4 property conversationally: users yesterday, top products, traffic patterns, no report builder involved. Once one first-party analytics platform does this, the rest of your vendor list follows or looks dated.
Second, and bigger: when the adtech industry built its own agentic transaction standard, the Ad Context Protocol launched in October 2025 by Yahoo, PubMatic, Scope3 and others, it was built on top of MCP. The protocol likely to carry agent-to-agent media buys chose MCP as its foundation. That tells you where integration budgets go next.
For a working stack, the practical change is the slow death of the point-to-point integration. My agency has years of accumulated glue: connector subscriptions, scheduled scripts, middleware that moves rows from ad platforms into dashboards. Each piece has its own auth, its own failure modes, its own maintenance bill. An MCP server collapses that pattern. The tool describes its capabilities once, and any agent I run, on any model, can use them. I run local open-weight models on my own hardware for parts of my marketing ops, and the same servers work there too, which is the point of a standard.
What I have learned building and reviewing MCP servers
I build MCP servers for marketing tools in my own automation practice, and I review them as an independent technical reviewer for AI adtech vendors. A few things the tutorials undersell.
Tool descriptions are the real API. The model decides which tool to call based on the plain-language description the server publishes. Vague descriptions produce agents that call the wrong tool with complete confidence. Early on, one of my own reporting agents kept pulling a misaligned date range because the tool description never stated which timezone the data was in. Boring documentation is now a functional requirement.
Scopes matter more than features. An MCP server that can read campaign data and also edit budgets is one prompt injection away from being a budget editor. When I review a server, the first thing I check is not what the tools do; it is what the credentials behind them are allowed to do. Read-only by default, mutations behind separate, explicit approval.
Latency compounds. A human tolerates a slow dashboard. An agent chaining five tool calls turns five slow seconds each into a 25-second reasoning step, and slow reasoning steps get skipped or guessed at. Fast, narrow tools beat slow, clever ones.
The spec still moves. MCP is young, and revisions land often enough that a server built in one quarter can behave differently against clients built the next. Treat an MCP integration like any other dependency in your stack: pin versions, test after upgrades, and budget a little maintenance. Less glue than the old world, not zero.
The question to ask your vendors
Ask every martech vendor on your list one question this quarter: do you have an MCP server, and what scopes does it require? The answer tells you whether they are building for the stack that is forming or the one that is fading, and the scopes tell you whether they thought about security for longer than a sprint.
Standards are boring right up until they are load-bearing. Programmatic buyers learned that with OpenRTB. MCP reached the load-bearing stage in about a year, mostly while marketing was watching the shinier parts of AI. The plumbing is the story.