This is the dated record of what is live on mcp.adtao.io, the tools your AI can use on your Google Ads accounts. The current release, v1.1 (July 2026), added the biggest lift: your AI can now check a Display placement or a search term against results from the whole network of connected accounts before you exclude or negate it, so it can warn you off blocking a term that earns money for advertisers like you. We add a new dated entry when something changes and never rewrite old ones, so you can always see exactly what shipped and when.
Sourced from the running server, so the version here matches what is deployed, not what is planned.
Current release
Released September 2026
v1.2·Sign in from Claude, approval codes, AI Max
Connect from Claude with your AdTAO sign-in, approve every change with a one-time code, and manage AI Max and experiments. Tool count went from 70 to 94.
Connecting and approving
→Connect from Claude on the web or Claude Desktop with your AdTAO email and password, no API key needed. The sign-in page asks whether your AI may propose changes; leave it unticked for a read-only connection.
→Every change that reaches Google Ads now needs a one-time code. Your AI proposes the change, you check it on the AdTAO approval page, and the page shows a code that works once, for 15 minutes, for exactly those changes. Your AI cannot get the code through AdTAO; only the signed-in person sees it. You can approve one change or several at once; saying no needs no code.
→Every plan, including free, has the same tools. Plans differ only in how many requests you can make.
New for your accounts
→A 0 to 100 health score with every issue found, and an audit of your conversion actions.
→Checks for budgets in the wrong places, bid targets, wasted spend by device, day and hour, and why traffic dropped.
→AI Max: whether it is on, which campaigns Google moved to it, where it sends traffic and what headlines it writes, and your search terms split by source. Changes to AI Max settings and AI Max experiments are proposed for your approval.
→A/B tests on Search campaigns (for example a new target CPA), proposed for your approval, with a readout as results come in.
→Negative keywords checked against comparable accounts before anything is called waste, with a watch list for terms that need more time, plus the negatives that block your own live keywords.
→Your brand searches versus competitors' searches, keyword gaps, ad copy ideas from competitors' ads, placement risk, and what Google is retiring in your account.
Clearer, safer answers
→Answers use Google Ads wording and plain English, and say plainly when something could not be read, including when a Google login has lost access to an account.
→Long answers keep their most important rows instead of being cut off, and query results include totals across every row read.
→Tool names now use only letters, numbers and underscores, as Claude requires. The old names still work.
→The privacy policy now says what connected accounts contribute to the shared data: a search pattern is shown only when at least three unrelated accounts share it, and a benchmark needs at least ten.
Retired
→Three tools that did not deliver what they promised: an insights tool that never returned results, a sync-notification tool that never sent notifications, and a theme recommendations tool that duplicated keyword gaps.
Released July 2026
v1.1·The intelligence layer
Added 24 new read tools, live on mcp.adtao.io. Your AI can now check Display placements and search terms against results from the whole network of connected accounts, not just your own. Tool count went from 46 to 70.
Display placement checks (20 tools)
→Before your Display ads run on a website, your AI can check it against 436,765 websites already scored for wasted-spend risk, using real placement spend from accounts like yours as the evidence.
→Ranked lists of the websites most likely to waste budget: odd domain endings, spammy-looking names, high-traffic drains, and high-CPC placements, each with per-website detail and a suggested exclusion.
→Reads your existing placement exclusions (shared exclusion lists, campaign links, account and campaign-level exclusions) and previews the exclusions it would add before anything changes.
→A read that scores how suitable a given website is for your account.
Search-term checks across the network (4 tools)
→Your AI can check a search term against real spend and conversions for that term across the whole network of connected accounts, so it can tell you whether a term earns money for advertisers like you before you touch it.
→Drill into how a term or phrase performs across accounts like yours, and whether it clears the bar for adding as a keyword or a negative.
Kept internal by design
→Applying exclusions, recomputing suitability, and running the network-wide checks stay internal for now. Nothing changes in a client account until the approval flow for these ships.
How this was checked
→Every tool was run against the live server and scored by pass/fail scripts, with no AI opinion involved. See the run from 12 July 2026: the intelligence layer, proven live.
Released June 2026
v1.0·Launch
The first public release of the AdTAO MCP server.
What your AI could do
→16 tools: 9 to read your account, 4 to stage changes, 3 for setup.
→Connects at https://mcp.adtao.io/mcp.
→Works in Claude Desktop, Claude Code (CLI), and Cursor.
What it could answer
→Account overview: list your accounts, headline performance, 12-week trends, and month-over-month comparison.
→Plain-English answers backed by your real account data, every figure traceable to a row in your own reports.
→A health scan that names the issue, how much wasted spend it costs, and the fix, across hundreds of checks.
→Live budget pacing across every account, so you can see which are over or under before the month closes.
→Every change staged for your approval first, one at a time or in bulk. Nothing changes in your account until you say yes, and the result is measured after.
→One approval at the manager (MCC) level flows down to every child account, where a manager hierarchy exists.
→Competitor and keyword-gap reads, so you can see the search terms you are missing.
Signing in
→An API key (adtao_...) that can only see your own account, from Settings, Security and Privacy.
→Rate limits by plan.
What it could not do yet
→No channel-level Performance Max exclusion, because the Google Ads API does not support it. Account and manager-level placement, URL, and brand exclusions are the levers you have.
→At launch we had published 5 of the 36 comparison tests we are building, so early results reflect those 5.
Ready-made skills
→Four to start: a weekly portfolio report, an account health scan, a budget pacing check, and safe bulk approval.
→Get them from the quickstart.
What is the latest release worth, and what should you do with it?
A tool that can only see your own account has no way to know, before you add a negative keyword or exclude a Display website, that the term or placement earns money for advertisers like you. That is the part v1.1 adds: your AI checks each proposed negative and each risky placement against results from the whole network of connected accounts first, which a setup limited to your own account cannot do. The everyday reads, the search terms report, the health scan, budget pacing, are things any good setup can do, and you should expect them to. The added value is the check that stops you blocking a term that quietly converts, and the 436,765 pre-scored websites that spare you building a wasted-spend list by hand.
What to do next: point your AI at your account, ask it to build a negative keyword list from your search terms report, and let it flag any term that earns for accounts like you before you paste the rest into a shared negative keyword list in the Google Ads UI. For Display, ask it to rank your current placements by wasted-spend risk and stage the exclusions for your approval. Nothing changes in your account until you approve it, so the first pass is a read you can act on within the hour.
How releases work
→Append-only. Past entries are never rewritten.
→Sourced from the running server. The version here matches what's deployed, not what's planned.