Paid media is the one MCP category with no official server in it. Google publishes one for Analytics and nothing for Google Ads; Meta publishes nothing at all. So the real choice is not which of a dozen options to pick — it is whether you run a community wrapper yourself or use a hosted product that holds the partner integration for you. This page is the paid-media half of the general MCP server roundup, at the depth that decision needs.
Last updated September 14, 2026.
The shape of this category
| Channel | Official MCP server? | What exists instead |
|---|---|---|
| Google Ads | No | Community wrappers; hosted products |
| Meta Ads | No | Community wrappers; hosted products |
| LinkedIn Ads | No | Hosted products only |
| TikTok Ads | No | Hosted products only |
| Reddit Ads | No | Hosted products only |
| Google Analytics | Yes — local, read-only, experimental | Google's own server |
| Search Console | No | Community wrappers |
That table is the argument. In every other MCP category the question is which vendor's server to add. Here, somebody has to build and maintain the integration, and the only question is who.
The options, ranked by what they actually do
1. Sprites: the channels behind one endpoint, with approvals
Ours, so weigh it accordingly. It is first because it is the only entry here that decides rather than executes: ask where Google Ads is wasting money and the server reads the live account, pulls search terms and change history, does the analysis, and returns changes with reasons attached. Google Ads, Meta, LinkedIn, TikTok and Reddit on one URL, through official partner integrations, with every write held as an approval card you edit before it applies.
Best for: anyone with production spend or a team. Watch for: it acts on channels you have connected inside Sprites; it is not a generic API wrapper you can point anywhere.
2. Community Google Ads wrappers
cohnen/mcp-google-ads and its forks expose a subset of the Google Ads API as MCP tools. For ad-hoc reporting they are genuinely good — "what was CPA by campaign last week" becomes a GAQL query the model writes and runs.
What it costs you: a developer token from API Center (basic access capped at 15,000 operations a day), a Cloud project, OAuth credentials and a refresh token, all in a config file on one machine. Where it ends: quota exhaustion mid-task as RESOURCE_EXHAUSTED, refresh tokens invalidated by a scope change, Performance Max asset-level operations usually missing, and no approval step anywhere. The full breakdown is on the Google Ads pillar.
3. Community Meta Ads wrappers
pipeboard/meta-ads-mcp is the common starting point. Read-only exploration works well; production does not. The Marketing API enforces per-user and per-app rate limits and every tool call is a fresh request, so an agent loop burns a day's quota in minutes and surfaces (#17) User request limit reached mid-task. User access tokens expire after 60 days, and nothing re-runs the consent flow for you.
4. Google Analytics: the one official server in the neighbourhood
Google's GA4 server covers account summaries, standard reports, funnels and realtime. Three limits Google states itself: experimental, read-only, and local — it runs on your machine and nowhere else. Useful beside an ads server, not a substitute for one.
5. Semrush: the competitive half of a paid-media question
Semrush's official remote server does not touch your ad account, and belongs here anyway: "who else is bidding on this and what are they saying" is a paid-media question, and this is the only official server that answers it. Projects access is deliberately read-only.
6. Zapier: actions, not analysis
Zapier MCP reaches ad platforms among its 9,000+ apps, so an assistant can pause a campaign or create an ad if that action exists. It cannot look at an account and tell you what is wrong. Billing is two tasks per tool call — the arithmetic is here.
7. n8n: the same, self-hosted and free
The MCP Server Trigger node turns a workflow into a callable server, and the Community Edition is free with unlimited executions. Same ceiling as Zapier — it runs the workflow you built — with the cost moved from a meter to your own operational time. Full comparison.
Hosted or self-hosted: the honest version
| Community wrapper | Hosted product | |
|---|---|---|
| Licence cost | Free | A subscription |
| Who gets the developer token | You, and its approval process | The vendor, as a partner |
| API quota | Your basic-access cap | Partner-grade |
| When the API ships a breaking change | You fix it | It is already fixed |
| Approval before a write | None unless you build it | A product feature, or not — ask |
| Audit trail across a team | None; it runs on one laptop | Shared, per change |
| Right for | One analyst, read-only, own account | Production spend, teams, multiple channels |
Both are legitimate. What is not legitimate is drifting from the first into the second without noticing — running a wrapper on a laptop against an account that has grown to real spend, with no approval gate and no shared log of what changed.
The question that matters more than the list
The MCP specification defines no approval primitive. A tool is a name, a description and a schema; calling it does whatever the server does. So on any server with write tools, this is in spec:
- You ask a loosely-worded question about a campaign that is underperforming.
- The model decides pausing it is the helpful next step.
- A pause tool exists and takes one argument.
- The campaign is paused. Nothing was exploited and nobody approved anything.
Whether a change is staged for a human is a product decision by the individual server, and it is invisible from a tool list. On a channel that can deliver a day's budget in an hour, ask it before you paste the URL. Is MCP safe covers the rest of the threat model.
Where to go next
- The general list across every marketing job: best MCP servers for marketers.
- On Claude specifically: the connector shortlist.
- The other facets: SEO and startup marketing.
- Setting one up: Claude, ChatGPT, or the whole hub.
MCP for advertising: common questions
Is there an official Google Ads MCP server?
No. Google publishes an official MCP server for Google Analytics and nothing for Google Ads. Every Google Ads MCP server in existence is either a community wrapper you host yourself or part of a commercial product.
Is there an official Meta Ads MCP server?
No. Meta publishes no MCP server for the Marketing API. The community wrappers that exist are unaffiliated projects built on the public API.
What is the best MCP server for advertising?
For anyone running production spend, a hosted server on official partner integrations with an approval gate on every write. For an analyst prototyping read-only reports against their own account, a community wrapper is a legitimate and free choice.
Can an MCP server change my campaigns?
If it exposes write tools, yes — and MCP defines no approval step, so a write reached by a model applies immediately unless the server chose to stage it. On a channel that can deliver a day’s budget in an hour, that distinction is the whole decision.
What do I need to run a Google Ads MCP server myself?
A developer token from API Center, a Google Cloud project with the Ads API enabled, OAuth credentials, and a refresh token you generate. Basic access is capped at 15,000 operations a day per token, and production access review expects a real product rather than an experiment.
Will connecting an agent to my ad account get it flagged?
Official API access through a partner integration will not. The risk sits with unofficial access patterns and with volume: an agent that fans out across every ad in every ad group looks nothing like a human, and quota exhaustion and fraud heuristics both notice.
Can Zapier or n8n run my ads through MCP?
They can execute individual actions you have configured, such as pausing a campaign. Neither reads the account and works out what should change, because that is analysis rather than automation — a workflow does what it was built to do.