5 Open-Source Google Ads AI Tools I Just Shipped — MCP, Gemini, Claude Skills, and the Full Stack

Live demo: MCP server, Gemini extension, and Claude Skills in action on a real Google Ads account.

The Problem That Wouldn't Go Away

I've managed Google Ads accounts for over 15 years — enterprise budgets, multi-location franchises, SaaS, ecommerce, lead gen. And for most of that time, "AI-powered Google Ads" has been a UX wrapper on top of the same manually-exported data we've always had. Fancier presentation. Same fundamental limitation.

The AI could describe data you'd already seen. It could not run a query. It could not check what's happening in your search terms right now. It could not pull budget pacing across your MCC at 9pm on a Thursday when something breaks.

The limitation was never the AI. The limitation was that nothing was actually connected.

Two things changed that. MCP (Model Context Protocol) became the standard for AI agents to use external tools — Claude, Cursor, Windsurf, OpenAI's Agents SDK all support it. And Gemini CLI launched an extension ecosystem. The infrastructure for building real AI tooling around Google Ads now exists.

So I built the tools. Five open-source projects that give AI agents genuine, live access to the Google Ads API. Not exports. Not screenshots. Real API calls, real data, your real account.

The Five Projects

And the fifth — Buddy at googleadsagent.ai — is the production system running on Cloudflare. Semantic memory, encrypted BYOK credentials, real-time WebSocket, monitoring, billing, and all the battle scars from running real accounts through it.

The MCP Server — The Foundation Layer

The google-ads-mcp server is the portable core. Install it once, and any MCP-compatible AI client gets live Google Ads access.

With this running, the AI doesn't approximate — it executes real GAQL queries against your account and returns actual data.

Every write tool is dry-run by default. Pass confirm=False first to preview what will change. Only confirm=True executes the mutation. This isn't a disclaimer — it's how the code is structured.

The Gemini CLI Extension — Batteries Included

The Gemini CLI extension is the opinionated, all-in-one package. It implements every feature type in the extension spec — the first Google Ads extension to do that.

You're prompted for credentials (stored in your system keychain, not plaintext). Everything else is pre-configured:

• 22 MCP tools — 15 read/audit + 7 write tools with confirmation gates
• Custom commands — /google-ads:analyze, /google-ads:audit, /google-ads:optimize
• Agent skills — PPC expertise with GAQL templates and anomaly detection
• Hooks — GAQL write blocker + audit trail logger
• Policies — Every API call requires explicit user confirmation
• Custom themes — Dark and light with Google's color palette

Claude Skills — Teaching Claude Google Ads

The newest addition. Google Ads Skills for Claude is an Anthropic Agent Skills package — 5 skills that teach Claude how to be a PPC specialist.

Three plugin bundles: google-ads-full (all 5), google-ads-readonly (no write access), and google-ads-mcp-only (just the server config). The wiki has full documentation across 11 pages.

I've submitted a Partner Skill PR to Anthropic's official skills repo to get listed alongside Notion.

Write Safety — The CEP Protocol

Every write operation across every project uses the same safety model: Confirm → Execute → Post-check.

1. Confirm — The AI shows exactly what will change (entity, current state, proposed state, risk level) and asks for explicit approval
2. Execute — Only after you say "confirm" does the mutation run
3. Post-check — The AI verifies the change took effect and notes how to reverse it

No write operation executes without your explicit approval. New campaigns are always created paused. Batch operations are itemized before confirmation. This protocol is enforced at the server level — not just in skill instructions.

How They Connect

Each project serves a different platform, but they share the same Google Ads expertise and safety protocols. Use whichever matches your AI client — or use all of them.

What This Means for Practitioners

The keyword era in paid media is ending. Not because keywords stopped mattering — they still route intent — but because the tactical layer of PPC management is being absorbed by automation. Smart Bidding, broad match, Performance Max, AI-generated ad copy. Google is systematically removing the levers that used to differentiate a good account manager from a mediocre one.

The practitioners who matter in three years are the ones who moved up the stack. Who understand the strategy above the automation. Who can structure business problems correctly and interpret what the machines are telling them.

But you can only move up the stack if the tactical layer is actually handled — not by a dashboard, but by an agent that has genuine API access, real data, and a safety model you trust.

That's what these five projects are. They're not demos. They're ported from production systems running real accounts. The GAQL templates, anomaly detection thresholds, write safety architecture — all from managing real budgets and fixing what broke.

Get Started