What is MCP (Model Context Protocol)? The standard that links AI to Google Ads
MCP is an open protocol that lets assistants such as Claude and ChatGPT call external tools. Its three components, how it works, hosted versus local servers, and its role in ad account management.
Tecrube 팀
마케팅 자동화
What MCP is
Model Context Protocol is an open standard that defines how an AI assistant talks to external systems. The assistant sees the tools an MCP server offers, for example “list campaigns” or “add negative keywords”, calls them when needed and brings the result back into the conversation.
The protocol was published by Anthropic in late 2024 and was quickly adopted by Claude, ChatGPT, Claude Code, Cursor and many other applications. The practical meaning of a standard is simple: a server written once works in every assistant that supports the protocol.
Why MCP was created
Before MCP, every link between an assistant and an external system was custom-built. A ChatGPT plugin for Google Ads here, a separate Claude integration there, each with its own maintenance. Every new assistant meant redoing the same work.
MCP reduced that link to one language. The ad platform side is a server, the assistant side is a client, and the contract between them is the protocol itself. That is why Tecrube's single connector URL behaves the same in Claude and in ChatGPT.
Three components: client, server, tools
The client is the application you chat in: Claude, ChatGPT, Cursor. The server is the service representing the external system, such as Tecrube's ads server. Tools are the operations the server advertises to the client; each has a name, a description and the parameters it expects.
When you say “list non-converting search terms”, the client picks the right tool, fills in the parameters (account, date range, threshold), sends the call to the server and shows you the returned table. The model decides; the server does the work.
How it works technically
When a client connects, it first requests the tool list and the server returns each tool with its description. Tools can also carry annotations such as readOnlyHint and destructiveHint, which the client uses to decide which calls need confirmation. From then on every call is a request-response pair: tool name and parameters go out, a structured result comes back.
In Tecrube every write tool is two-phased: the first call only produces a preview, the second carries the approval with the preview's identifier. That is the protocol-level form of the preview → approve → apply flow.
Hosted versus local servers
A local server runs on your own computer; setting it up requires a terminal, API keys and a Google Cloud project, and you handle updates yourself. It suits developers and data teams.
A hosted (remote) server runs on the internet; you add only a URL to the client, authorisation happens through an OAuth screen and maintenance belongs to the server's operator. Tecrube is the hosted model: setup takes minutes and the same URL works in the Claude app on your phone.
Managing ad accounts with MCP
Tecrube's server exposes Google, Meta, Microsoft, TikTok and LinkedIn ad accounts, plus Google tools such as Search Console and Tag Manager, through one connector. Read tools answer immediately; write tools return a preview and wait for approval; a single-step budget increase above fifty percent is rejected; every approval is logged.
Setup takes three steps: sign up for Tecrube and copy the connector URL from your member panel (/app); add it to Claude's or ChatGPT's connector settings; on the first tool call, pick your ad account on the platform's OAuth screen. Sign-up includes 75 credits that never expire; afterwards Pro is $49 a month, Agency $99 a month and Enterprise starts at $299.
Which applications support MCP
On desktop and web, Claude and ChatGPT let you add a remote MCP server from their connector settings. Among developer tools, Claude Code and Cursor recognise the same server through a configuration entry; one command in Claude Code, one line in Cursor's MCP settings. On mobile, the Claude app automatically sees the connectors attached to your account.
The practical result: the marketing team chats in Claude or ChatGPT, a developer wires the same ad tools into a script in Claude Code, and the agency owner asks for the morning report from a phone. All three use the same Tecrube URL and the same approval flow.
Security: how access is limited
There are three boundaries. The authorisation boundary: only the accounts you select in OAuth are visible, and no password is shared. The tool boundary: the server offers only the tools it defines, so the assistant cannot do arbitrary things in the account. The approval boundary: write tools are never applied without a preview.
Those three boundaries turn the worry of “I am opening my account to an AI” into the reality of “I am letting an AI use specific tools, with my approval”.