Best AI Agent Skills for Google Ads & Meta Ads (2026)

An AI agent with your Google Ads credentials can spend money. That single sentence should change how you evaluate every skill in this category, and it's the reason this list is shorter and more paranoid than most.

Ad platform skills are among the most useful things you can connect to Claude Code. They're also the only category on this site where a bad install has a direct, immediate financial cost. Below are the four real options in our database, what each one covers, and the access model I'd insist on before connecting any of them.

The four skills

Every one is free, and every one is community tier. There is no official or verified ad platform skill in the directory today. That absence is the single most important fact on this page. Google and Meta both publish official APIs, and nobody has shipped a first-party agent skill on top of them. You are choosing between independent implementations of somebody else's API.

meta-ads-mcp

Install: claude mcp add meta-ads-mcp -- npx -y pipeboard-co/meta-ads-mcp

The most-starred option at 1,131 stars, last updated 2026-07-30, covering both Facebook and Instagram Ads. It's categorized for Coding and Marketing and lists compatibility with Claude Code, Cursor, and VS Code Copilot, which is wider agent support than the others here.

Use it when Meta is your primary channel and you want campaign management and reporting in the same session where you're drafting creative. The active maintenance is the strongest signal in its favor. An ad platform integration that goes six months without an update is a liability, because Meta changes its API on its own schedule.

google-meta-ads-ga4-mcp

Install: claude mcp add google-meta-ads-ga4-mcp -- npx -y irinabuht12-oss/google-meta-ads-ga4-mcp

The broadest scope in the directory: Google Ads, Meta Ads, and GA4 in one server, advertising 250+ tools for campaign management, analytics, and optimization, with compatibility across Claude Code, Cursor, Codex CLI, and VS Code Copilot. 1,049 stars.

The appeal is obvious. One install covers spend and outcomes, so an agent can trace a campaign to conversions without you switching context. Two things temper it. Its last update is 2026-04-07, roughly four months stale as of this writing, which matters more for ad APIs than for most software. And a 250-tool surface is a lot of capability to hand a single credential set.

I'd pick this one specifically when cross-channel reporting is the job. For pure campaign execution on one platform, a narrower skill is the safer buy.

mcp-google-ads

Install: claude mcp add mcp-google-ads -- npx -y cohnen/mcp-google-ads

Connects Google Ads to Claude and other agents for natural-language analysis of campaign information, performance metrics, and keyword analytics. 673 stars, and the oldest entry here at last update 2025-10-16.

That date is the problem. Google Ads API versions deprecate on a published schedule, and a connector nearly a year without an update is one deprecation away from breaking. Read the repo's recent issues before you install. If the maintainer is responsive there, the stale timestamp matters less than it looks.

Its framing is analysis-first rather than management-first, which I consider a feature. Reading campaign data is the lower-risk half of this category and where most of the value is anyway.

facebook-ads-mcp-server

Install: claude mcp add facebook-ads-mcp-server -- npx -y gomarble-ai/facebook-ads-mcp-server

349 stars, updated 2026-07-29, Claude Code only. Its directory description is generic, which usually means the metadata was auto-generated rather than written by the author.

It's the smallest project of the four and I have the least to say about it. Recent maintenance is a point in its favor. If meta-ads-mcp doesn't fit your setup, this is the alternative on the same platform.

The access model, which matters more than the skill

Here is the part I'd care about more than any feature comparison.

Create a dedicated user with read-only access first. Both Google Ads and Meta Business Manager support granular permissions. Give the agent a login that can read reports and cannot edit campaigns, budgets, or bids. Run it that way for a few weeks. Most of what you want from these skills, including performance analysis, keyword mining, and cross-channel attribution, needs nothing more.

Grant write access per-account, never at the manager level. An MCC or Business Manager token reaches every client account under it. If you run ads for clients, a compromised manager credential is not your incident, it's theirs.

Set account-level spend caps in the platform, not in the agent. The cap that protects you is the one Google or Meta enforces. An agent asked to respect a budget is following an instruction; a platform-level cap is a wall.

Read the code before installing. All four repos are public. You're looking for what the server does with credentials, whether it phones anywhere other than the ad platform, and whether write operations are gated. This takes twenty minutes and it's the only real defense at community tier. Our guide to what can go wrong with untrusted skills covers the failure patterns, and why author trust tiers matter explains what the tier labels actually promise.

None of this is unique to ad skills. It's just that here the blast radius is denominated in dollars per hour.

A workflow that survives contact with a real account

The pattern I'd recommend is a weekly read-only audit rather than a live agent sitting on your account.

Once a week, in a fresh session, ask the agent for four things: search terms from the last seven days with spend above your click cost but zero conversions, ad groups where impression share loss to budget exceeded 20%, any campaign whose cost per conversion moved more than 30% week over week, and creative or keywords that stopped serving entirely. Those four queries catch most of what goes wrong in a paid account, and none of them require write permission.

Then apply the changes yourself in the platform UI. This sounds like it throws away the automation, and it partly does. What it buys you is a reviewable diff between what the agent recommended and what you actually did, which is the record you'll want when a campaign turns and you're reconstructing why.

Two habits make the weekly audit noticeably better. Feed the agent your conversion definitions at the start of the session, because a model reading raw campaign data has no idea whether your "conversion" is a purchase or a newsletter signup, and the recommendations diverge sharply depending on which. And give it the account's recent history in one line: what launched, what you paused, what seasonality you expect. Most bad agent recommendations in paid media come from missing context rather than bad reasoning, and one sentence of framing fixes a surprising share of them.

Scale up from there only if the recommendations hold up. If three consecutive weeks of suggestions would have been right, you have a case for granting narrow write access on one campaign. If they wouldn't have been, you learned that for free.

What these skills do well

Pulling a month of search terms and clustering them by intent, a task that's tedious in the Google Ads UI and trivial for an agent with API access.

Comparing spend against outcomes across Google and Meta without exporting three CSVs first.

Drafting negative keyword lists from actual search term data rather than from imagination.

Answering questions the platform UIs make hard, like which ad groups have impression share loss concentrated in specific hours.

What they're bad at is judgment. An agent will happily tell you a campaign is underperforming without knowing you launched it four days ago, or recommend pausing a keyword that assists conversions it can't see. Treat the output as a fast analyst's first draft.

For the broader marketing stack, see our list for marketers and our roundup of agent skills for marketers. Social-side practitioners will find more in tools for social media managers.

FAQ

Is there an official Google or Meta agent skill? No. As of August 2026 every ad platform skill in our directory is community tier. Both companies publish official APIs; neither has published a first-party agent integration.

Can these skills change my campaigns, or only read them? It depends on the skill and, more importantly, on the credentials you supply. Several advertise campaign management capability. The reliable control is platform-side permissions. Give the agent a read-only user and write access is impossible regardless of what the skill supports.

Which one should I install first? meta-ads-mcp if you're Meta-heavy, mcp-google-ads if you're Google-heavy and mainly want analysis, google-meta-ads-ga4-mcp if you need both channels plus GA4 in one place and can accept a staler codebase.

Do MCP servers work with agents other than Claude Code? Some do. meta-ads-mcp and google-meta-ads-ga4-mcp list Cursor and VS Code Copilot compatibility; google-meta-ads-ga4-mcp adds Codex CLI. mcp-google-ads and facebook-ads-mcp-server are listed for Claude Code only.

What's the realistic failure mode? Not a dramatic hack. It's an agent with write access making a well-intentioned bulk change, pausing a set of keywords or adjusting bids across an account, that you didn't review and can't easily reverse. Read-only access removes this entirely.