← All posts

MCP Servers for Ads: What Claude and ChatGPT Can Actually Do With One

Nov 15, 2026

MCP Servers for Ads: What Claude and ChatGPT Can Actually Do With One

The piece most people are missing about "AI running ads"

Everyone keeps asking whether Claude or ChatGPT can run their Google or Meta ads. The honest answer is: not by themselves. A language model on its own cannot see your campaigns, pull yesterday's spend, or pause a losing ad set. It needs a bridge. That bridge is called an MCP server.

If you've been hearing the term and nodding politely, this post is the one to bookmark.

What MCP actually is

MCP stands for Model Context Protocol. It's an open standard, introduced by Anthropic in late 2024 and now supported by ChatGPT, Claude, Cursor, and a growing list of other clients. Think of it as USB-C for AI assistants: a single, predictable way for a model to connect to outside tools and data.

An MCP server is just a program that exposes a set of "tools" (functions the model can call), "resources" (data the model can read), and "prompts" (pre-written workflows) using that protocol. It runs wherever you want — on your laptop, in a container, behind your company's firewall — and the AI client talks to it over a defined channel.

The model doesn't magically know what the server can do. The server advertises its capabilities on connection, and the model picks from that menu.

What an MCP server "for ads" specifically does

An ads-focused MCP server wraps the Google Ads API, Meta Marketing API, LinkedIn, TikTok, or whatever platforms you care about, and exposes them as tools a model can call in plain language.

A well-designed one typically offers tools along these lines:

  • `get_account_summary(account_id, date_range)`
  • `list_campaigns(account_id, status)`
  • `get_search_terms(campaign_id, date_range)`
  • `get_ad_performance(ad_group_id, metrics)`
  • `add_negative_keyword(campaign_id, term, match_type)`
  • `pause_ad(ad_id)`
  • `adjust_budget(campaign_id, new_daily_budget)`
  • `create_ad_variant(ad_group_id, headlines, descriptions)`

Each tool has a schema — required fields, allowed values, return format. The model fills in arguments based on your conversation and gets structured JSON back.

That's it. No mystery. The "AI" part is the model deciding which tool to call, when, and how to interpret the result.

What Claude or ChatGPT can do once connected

With an ads MCP server attached, here are the kinds of jobs that go from impossible to trivial in a chat window:

Investigation. "Why did CPL double last week?" The model pulls daily performance, compares week-over-week, checks search terms for junk, looks at impression share lost to budget, and writes up what it found. In one message.

Waste removal. "Show me every search term from the last 30 days that cost more than $20 with zero conversions, and draft a negative keyword list grouped by theme." You review, approve, and the model applies them.

Creative iteration. "Pull the top three performing responsive search ads in the Dallas campaign, identify the headline patterns that are winning, and draft four new variants for me to review."

Budget shifts. "Rank my campaigns by cost per qualified lead this month and suggest a reallocation if I keep the total daily budget flat."

Routine reporting. "Every Monday at 8am, summarize last week's spend, conversions, CAC, and top three issues across both ad accounts."

None of this requires custom code on your end. It's a conversation with a model that happens to have hands.

What it can't do, and shouldn't

An MCP server is only as capable as the tools its author built. If `pause_campaign` isn't exposed, the model can't pause campaigns — no matter how cleverly you ask.

That's a feature, not a bug. A thoughtful ads MCP server deliberately withholds high-risk operations, or puts them behind a human confirmation step. You do not want a model deleting ad groups because it misread a request. The protocol supports "tool call approval" flows for exactly this reason.

It also doesn't give the model memory between sessions unless you connect it to something that does — a CRM, a notes database, a logging server. On its own, each conversation starts fresh.

Where this gets interesting for a small business

The single-chat use case is already useful. The bigger shift is when the same MCP server powers an always-on agent instead of a human typing prompts.

That's the loop we run inside Plyto: one set of tools, called by an agent on a schedule, feeding results back into the CRM so attribution, nurture, and spend decisions share the same brain. The ads MCP layer is what lets the agent actually touch Google and Meta instead of just talking about them.

If you're building this yourself, a reasonable first project is a read-only MCP server that exposes performance queries to Claude Desktop or ChatGPT. You'll get 80% of the analytical value with none of the risk. Add write tools — negative keywords, budget nudges, ad pauses — one at a time, each behind a confirmation prompt, as you build trust in how the model uses them.

The short version

An MCP server for ads is a small piece of software that turns "ChatGPT, look at my Google Ads account" from a party trick into a real workflow. The model brings the reasoning. The server brings the hands. What you get depends entirely on which tools the server exposes and how carefully they were designed.

That design choice — which tools, which guardrails, which data flows back into the CRM — is where the actual product lives. Everything else is protocol.

Plyto is your AI marketing assistant: it makes and runs your ads

It writes the ad, builds the landing page, captures the leads and shows you which dollar came back. Try it free, no credit card.

Start free