AI Agents for Facebook & Google Ads: MCP Servers & CLIs in 2026

In 2026 you can point an AI assistant at your ad account and manage it in plain English — no Graph API calls, no GAQL. The bridge that makes this work is MCP, the Model Context Protocol. MCP for ads has gone from a hobbyist experiment to something both Google and Meta now ship officially. Here’s what actually exists — the official and community MCP servers, the CLIs, and the aggregator platforms for Facebook and Google Ads — what each can and can’t do, and the one safety rule that matters before you connect an autonomous agent to a live budget.

The single most important fact up front: Google’s official ad MCP is read-only; Meta’s official MCP can write. That difference drives everything below.

Quick navigation

• What is MCP (and why ad managers care)
• Official MCP servers: Google vs. Meta
• Community MCP servers & CLIs
• Aggregator & agent platforms
• What you can (and can’t) do today
• Is it safe to connect AI to your ad account?
• How to connect one to Claude
• The near-term outlook
• FAQ
• Bottom line

What is MCP — and why ad managers suddenly care

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, that lets AI assistants connect to external tools and data through one uniform interface. Anthropic’s own analogy is that MCP is “a USB-C port for AI applications.” An MCP server wraps an external system — a database, an API, an ad account — and exposes it as a set of typed tools the AI can call; an MCP client (Claude Desktop, Claude Code, Cursor, ChatGPT, Perplexity) discovers and invokes those tools from a natural-language request.

For advertising, an ad-platform MCP server sits between your AI assistant and the platform’s marketing API (the Meta Marketing API or the Google Ads API). Instead of writing a query, you ask — “show me last week’s ROAS by campaign” or “which ad sets are under 1% CTR?” — and the server translates that into the right API calls and returns structured results. This is the same idea behind the custom Claude Skills we build for repeatable work, extended to a live ad account.

Official MCP servers in 2026: Google vs. Meta

The big 2026 shift is that the platforms themselves entered the game — and they made opposite choices about how much power to hand an AI.

The core contrast: Google open-sourced an official Google Ads MCP server that is strictly read-only — it cannot change bids, pause campaigns, or create assets. Meta shipped official “Ads AI Connectors” (an MCP server and a CLI) that are write-capable — you can create and edit campaigns, ad sets, and ads in natural language. Same protocol, two very different risk postures.

Google Ads MCP (official, read-only). Google open-sourced its Google Ads API MCP server in October 2025 (repo: googleads/google-ads-mcp). It exposes just three tools — list your accessible accounts, run a GAQL query to pull metrics/budgets/status, and fetch resource metadata — and Google’s docs are blunt that “this implementation is strictly read-only. It cannot modify bids, pause campaigns, or create new assets.” Google has said future versions could expand beyond read-only. It runs locally or on Cloud Run. (This is the exact server connected in the environment we used to research this piece.)

Meta Ads AI Connectors (official, write-capable). Meta launched its Ads AI Connectors in open beta in April 2026 — an official MCP server and an official CLI as one product family. Meta groups the capabilities into four areas: comprehensive reporting, campaign management, catalog management, and signal diagnostics — and explicitly advertises that you can “create and edit ads, ad sets, and campaigns using natural language.” Meta positions it as no-code (Business login and scope approval rather than developer credentials).

Community MCP servers & CLIs

Before the platforms shipped their own, the community built the category — and those projects are still the most popular way in, especially for writes on Google (where the official server won’t).

Meta / Facebook Ads

• pipeboard-co/meta-ads-mcp — the market leader, with just over 1,000 GitHub stars as of mid-2026 and active development. It’s write-capable with built-in guardrails (writes are explicit, new campaigns start paused where the platform allows), and it ships as both a hosted remote server and a self-hostable one, with a free tier plus paid plans.
• Several smaller TypeScript/Python servers (brijr/meta-mcp and others) wrap the same Marketing API with varying tool counts; most are early-stage — check stars and last-commit date before trusting one with a budget.

Google Ads

• cohnen/mcp-google-ads — the popular original community server (~650 stars), read-only like Google’s official one: list accounts, run GAQL, pull campaign and ad performance. MIT-licensed, works with Claude Desktop and Cursor via OAuth or a service account.
• A handful of newer community servers add analysis/optimization framing on top, but the official write path for Google Ads still runs through the SDK, not Google’s MCP — the few community servers that do write are early-stage and wrap the same SDK.

The CLIs and SDKs underneath

• Meta Ads CLI (official, launched April 2026) — full campaign lifecycle from the command line, engineered for agents and CI: JSON/table/plain output, environment-variable secrets, defined exit codes. Python 3.12+.
• google-ads (official Python client) and facebook-business (Meta’s Business SDK) are the full read/write foundations that almost every server above is built on. If you need Google Ads writes today, this is the layer that does it.

Aggregator & agent platforms

If you’d rather not self-host, several platforms expose ad APIs to agents as managed tools — handling OAuth, token refresh, and multi-account access for you:

• Composio — hosted Google Ads and Meta Ads MCP toolkits with managed OAuth and enterprise security (SOC 2 / ISO 27001), plus integrations for Claude Code, the Claude Agent SDK, and other agent frameworks.
• Adspirer — an “AI advertising agent” exposing 300+ tools across six ad platforms (Google, Meta, LinkedIn, TikTok, and more) plus analytics.
• Zapier MCP, Pipedream, and n8n — workflow platforms that reach Google and Meta ads through their connectors; strong for triggers, leads, and reporting pipelines, lighter on full campaign management. The n8n ad-ops patterns are notable for baking in human approvals and audit trails.

What you can (and can’t) do today

Works today:

• Read and analyze in natural language — spend, impressions, CTR, CPC, conversions, ROAS by campaign/ad set/ad; arbitrary GAQL on Google; search-term and keyword performance; account diagnostics. Every server above does this.
• Create and edit campaigns by voice on Meta — via Meta’s official connectors or Pipeboard.
• Write to Google Ads programmatically — but only through the SDK, not Google’s official MCP.
• Generate reports and pipe them into other tools through the CLIs and workflow platforms.

Not yet (through the official open surfaces):

• Autonomous bid/budget changes on Google Ads via the official MCP — blocked by design (read-only).
• True “AI runs the whole campaign,” hands-off — that’s a roadmap ambition, not a shipped MCP capability (see the outlook).
• One agent that writes to both Meta and Google out of the box — you have to stitch platforms together (Composio, Adspirer, n8n); no single official server spans both.

Is it safe to connect an AI agent to your ad account?

This is the question that matters, because the thing on the other end of the connection spends money. The failure modes are real: a hallucinated or fat-fingered write (a $500/day budget where you meant $50, a bid 10× too high), pausing the wrong campaign, a bad date filter driving a wrong “optimization,” and — critically — no undo for spend already incurred. There’s also prompt-injection risk: an agent that reads your account could act on an instruction hidden in a campaign name or comment.

The strongest signal that the industry takes this seriously is that Google deliberately shipped its official MCP read-only. When the platform that owns the API decides autonomous writes aren’t ready, that’s your cue too.

The guardrails to insist on — most of which already ship — reflect the same discipline we apply to every ad account we touch:

• Read with anything; write behind a human. Use read-only servers freely for analysis. Never let an agent apply a budget, bid, or status change without explicit human approval.
• Paused-by-default creation — new campaigns should start paused (Pipeboard does this) so nothing spends before you review it.
• Scoped, revocable credentials — least-privilege OAuth, per-environment tokens you can revoke (managed platforms like Composio help here).
• Start on a test or low-budget account before you point an agent at your main spend.

If AI-driven changes ever get an ad account flagged, the recovery playbook is the same as always — see our guide to AI tools for Facebook ads for the account-safety section, and keep a human in the approval loop.

How to connect an ad MCP to Claude (the short version)

1. Pick a server by risk. For read-only analysis, the official googleads/google-ads-mcp or cohnen/mcp-google-ads. For Meta writes, Meta’s official AI Connectors or Pipeboard.
2. Get credentials. Google needs a developer token + OAuth (or a service account); Meta’s connectors use Business login and scope approval — no developer credentials required.
3. Register the server in your MCP client (Claude Desktop/Code config, Cursor, etc.) — locally, or use a hosted/managed option to skip setup.
4. Start read-only, confirm the tools work with a few reporting prompts, and only then consider a write-capable server — with approvals on.

The near-term outlook: hype vs. shipped

Both platforms are steering toward “agentic” advertising, but keep the roadmap and the reality separate.

Meta’s stated vision (first reported by the Wall Street Journal in June 2025): by the end of 2026, a business supplies a product image, a budget, and a goal, and Meta’s AI generates the creative, picks the audience, sets and optimizes bids, and measures — with no human intervention. Treat this as Meta’s ambition. What actually shipped in April 2026 was the AI Connectors — assistive tools you drive, not an autonomous campaign-runner.

Google used Google Marketing Live 2026 to push its agentic direction — “Ask Advisor,” a Gemini-based agent spanning Google Ads, Analytics, and Merchant Center, framed as an assistive strategic partner rather than a fully autonomous one. The read-only official MCP is Google’s cautious developer track running alongside those first-party features.

And the useful counterweight to the hype: Gartner predicts that over 40% of agentic-AI projects will be canceled by the end of 2027, citing cost, unclear value, and weak risk controls. The winners won’t be the teams that handed everything to an agent — they’ll be the ones who kept a human on the money.

Frequently asked questions

Bottom line

MCP has turned “talk to your ad account” from a demo into something you can actually run in 2026 — with official servers from both Google and Meta and a healthy community layer. Start where the risk is lowest: connect a read-only Google Ads MCP to Claude and let it do your reporting and analysis today. Add write-capable Meta connectors only with human approval in the loop, a paused-by-default setup, and a test budget. The platforms are racing toward autonomous, agent-run campaigns — but the teams that win the next year will be the ones who let AI read and draft while a human still signs off on the spend. For the tools that do the creative and optimization work around this, see our guides to the best AI tools for Facebook ads and the best AI tools for Google Ads in 2026.