5 Open-Source Google Ads AI Tools I Just Shipped

Source: YouTube auto-generated captions, lightly cleaned (repeated caption lines removed, misheard brand names corrected).

0:00 — A third of spend bought nothing

This morning, I pulled the last 30 days from one live account under management. Total spend was $939. $325 went to search terms with zero conversions. That is about a third of the budget buying nothing. Average CPA sat at $41 and 101 distinct terms were already negative keyword candidates waiting on a decision. Those 101 terms are not a research project. They are concrete negative keyword candidates sitting in the search terms report right now.

Most teams still export the CSV, filter in a sheet, and hope someone has an hour free. That lag is where the same zero conversion queries keep charging overnight while average CPA holds near $41.

0:49 — Twenty-nine tools, dry-run first

I just shipped a set of open-source Google Ads AI tools to close that lag. The package exposes 29 tools an agent can call. Every right path defaults to dry run. So, the first pass shows the planned change without touching the live account. You review the plan, then you allow the mutate. That is the whole safety model in one sentence. The stack is not five disconnected demos. It is one path from local agent runtime to Google Ads with a hard confirm gate. You stand up the MCP server first, then the Gemini extension, then the Claude skills that teach the agent how media buyers actually work. Only after that do you open the mutate tools, still behind dry run and confirm before mutate.

Step one is the MCP server. MCP is the

1:37 — Step 1 — Open the MCP server

protocol that lets a desktop agent call structured tools instead of pasting screenshots into chat. You run the server beside your Ads workflow, so the agent can list campaigns, pull search terms, and draft negatives through named functions. Without this layer, the model is guessing from text. With it, the model is calling the same operations you would click. You install the MCP server before any model specific piece because every later skill assumes those tool names exist. If Gemini or Claude loads first, the agent invents tool shapes that do not match production. Starting with MCP locks the contract. Search terms, negative drafts, bid reads, and change previews all share one schema the rest of the stack can trust. Step two is the Gemini extension. This wires the same

2:28 — Step 2 — Add the Gemini extension

MCP tools into Gemini, so you can work from the surface you already use for long context. The extension does not replace the server. It registers the 29 tools inside Gemini and keeps dry run as the default right mode. You ask in plain language, the extension roots to MCP, and you get a structured plan back instead of a paragraph of advice. Gemini earns its slot here because search term reviews and negative clustering benefit from a wide context window.

You can drop a full terms export summary into the thread and still keep the tool calls tight. The extension matters only after MCP is healthy. Otherwise, Gemini has nothing reliable to call, and you are back to copy-paste analysis that never reaches the account. Step three is the Claude skills pack. Skills are short

3:18 — Step 3 — Load Claude skills

instruction modules that teach the agent how a media buyer thinks about waste, match types, and shared negatives. They sit on top of the MCP tools, so Claude does not only fetch rows, it groups the 101 candidates, ranks by spend, and proposes a negative list you can accept or edit before anything is written. Skills come after Gemini because they encode judgment, not transport. A skill can say ignore brand terms, prefer exact match negatives first, and cap the first pass at the highest spend zeros. That guidance is useless if the tool layer is missing.

Loaded in this order, Claude becomes a buyer assistant on the same 29 tools instead of a generic chatbot describing Google Ads theory. Step [snorts] four is the dry run default on every right tool.

4:08 — Step 4 — Default to dry-run

When the agent wants to add negatives or adjust a bid, the tool returns the planned payload and stops. You see the 37 or 101 changes as a preview table, not as a surprise in the change history. Live mutate stays locked until you deliberately switch modes. That single default is what makes open source agents safe on real spend. Dry run is a product decision, not a nice to have flag. Agents will eventually propose a bulk negative list against a $325 waste pocket.

If the first call writes live, one bad cluster can pause good traffic. Returning a plan keeps the human in the loop at the exact moment cost appears. You approve the plan because the numbers justify it, not because the model sounded confident. Step five is confirm before mutate. Even after you leave dry run, the mutate tools demand an explicit

5:02 — Step 5 — Confirm before mutate

confirm token on the next call. That second gate stops a runaway loop from applying the same change twice and stops a partial plan from shipping because a chat window scrolled. The agent must restate what it will change, then you confirm, then the right executes once. Confirm before mutate sits last because it only has meaning when the earlier layers already produce a precise plan. You are not confirming a vague idea. You are confirming a dry run payload that lists campaign, ad group, keyword text, and match type. On the account from this morning, that means reviewing which of the 101 terms enter the negative list before a single one is applied.

This order matters because each layer removes a different failure mode. MCP removes invented tool names. Gemini and

5:54 — Why this order matters

Claude skills remove generic advice that never becomes an operation. Dry run removes silent live rights. Confirm before mutate removes accidental double applies. Skip ahead and you either get a chatbot with no hands or an agent with hands and no brakes on a $939 monthly account. Anchor the workflow on the live numbers, not on a demo account. Pull 30 days, rank zero conversion terms by spend, and let the skills propose negatives for the $325 pocket first.

Run that proposal through dry run. You should see the candidate count, the matched campaigns, and the no OP guarantee before confirm. That is how 29 tools stay useful on a morning when average CPA is $41. Recap the path in plain order. Stand up the MCP server so the agent has real Google Ads tools. Add the Gemini extension for wide context review.

6:52 — Recap

Load Claude skills so ranking and negative strategy match how buyers work. Keep every right on dry run by default. Finish with confirm before mutate so nothing hits the account until you say so. Five open pieces, one safe loop. On the sanitized account from this morning, that loop targets $325 of zero conversion spend inside a $939 month without guessing. The full write-up of the 29 tools, the dry run defaults, and the confirm date is on the blog. The blog is on ahmeego.com.

Buddy by Ahmeego, a company of Bit All Started with an idea.