For digital marketers, the acronym MCP now matters almost as much as ROAS. The Model Context Protocol is the open standard that lets AI assistants discover and call outside tools — including the APIs that create, read, and change paid campaigns.
Anthropic published MCP in November 2024. By December 2025 it had been donated to the Agentic AI Foundation under the Linux Foundation, with more than 10,000 active public servers. It is now the connective layer behind a fast-growing list of advertising platforms.
What MCP actually does
MCP standardizes how an AI application talks to a server that exposes capabilities. Instead of wiring every model to every data source, each client and server implements the protocol once. Servers can expose tools for executable actions, resources for read-only context, and prompts for reusable templates.
Messages travel over JSON-RPC 2.0. Local processes use stdio; remote endpoints use Streamable HTTP. Authorization is built on OAuth 2.1, which matters because an agent should only see the accounts its signed-in user can already open.
From read-only reports to live campaign edits
Ad platforms adopted MCP quickly but unevenly. Google’s open-source Ads API server arrived in October 2025 with read-only access and just three tools. Amazon Ads opened a beta in November 2025. Meta changed the stakes in April 2026 by letting Claude and ChatGPT create and edit live campaigns, then opened the same server to any developer with a Meta app in July 2026.
TikTok, X, Microsoft, and Adform have followed with very different permissions. X exposes 23 tools, ten of them write-capable. Adform published 29 skills that are deliberately read-only. The same protocol can carry reporting or budget changes — platform policy decides which one you get.
What performance teams should check
Before connecting an AI assistant to a paid platform, run a quick governance audit.
- Read vs write: Know whether a server can only pull reports or can pause campaigns and change budgets.
- Auth scope: Confirm the agent uses the signed-in user’s OAuth token, not a broader service account.
- Tool poisoning risk: Tool descriptions and outputs both enter model context. Treat external inputs as untrusted.
- Token cost: Stacked MCP connections multiply context loads. Consolidate calls where possible.
The July 28, 2026 revision removed protocol-level sessions and the initialize handshake. For ad tech vendors, that forced infrastructure rebuilds. For marketers, it signals the protocol is maturing fast and should be tracked like a product, not a passing trend.
Don’t confuse MCP with AdCP, the Ad Context Protocol built on top of it for advertising tasks, or with Google’s A2A, which standardizes agent-to-agent communication. MCP is the lower layer that standardizes how an agent talks to a tool.
Source: PPC Land
