I build iOS apps almost entirely with AI agents. The code, the App Store screenshots, the store listing, most of it runs without me touching much. One step never got automated, and it's the one that decides whether any of it makes money: marketing.
I wasn't stuck on running ads. I was stuck earlier than that. I didn't know what my competitors were advertising, so I had nothing to base a creative on. My whole method was opening a competitor's Tiktok, finding a video with good view counts, and telling my agent to make something in that format.
AdKit is a tool that plugs into Claude or ChatGPT and both answers that question and runs the ads. I tested it on PairCare, my own birth control app for couples, which I'd just rebranded around a new feature. I wanted to know whether the rebrand landed. Four days and about fourteen dollars later I had an answer, and it wasn't the one I expected.
What Is a Meta Ads MCP?
MCP is the standard that lets a tool hand its functions to an AI assistant. A Meta Ads MCP means Claude or ChatGPT can call the Facebook Ads API directly, reading your campaigns and in most cases changing them, from an ordinary chat window.
People connect one for three reasons, roughly. To ask questions instead of building reports. To combine Meta numbers with Google, GA4 or Shopify in a single question. Or, like me, to avoid learning Ads Manager in the first place.
That third one is worth explaining, because it sounds lazy and isn't. Ads Manager asks you to pick an objective, then build three nested layers of campaign, ad set and ad, then choose a bidding strategy, placements, an optimisation event and a budget structure, all before a single ad goes live. Most of those choices only make sense once you understand how delivery works, which you learn by getting them wrong. Typing "target the US, put it on Instagram" skips that.
Mine was the third reason, and it mostly held up. I barely opened a dashboard during this test, and every number in this review came back through chat, including the daily breakdown I use later to explain what went wrong.
What it doesn't mean is that you never open Ads Manager again. Account setup is still manual, and when Meta rejected one of my ads, that rejection never reached the agent. More on both below.
See What Your Competitors Are Advertising Right Now
This is the part that solved my actual problem, so I'll start here.
I gave it the names of five competing apps, Flo, Clue, Natural Cycles, Stardust and Paired, and it came back with every ad those apps are currently running. Not a summary of them but the actual video file, the headline, the full body copy, and the destination URL.
That last part matters more than it sounds. The video ads come with a mediaUrl you can download the real mp4 from. So the research doesn't stop at "here's what they're saying." You end up holding the file.
The dashboard splits the same data into tabs: creatives, copy, headlines, link descriptions, and the landing pages competitors send people to. Each entry carries how long it's been running and how many ads it's used in.
The Copy tab gives you full body text, not excerpts.
You can do part of this for free, and it's worth being clear about which part. Meta's ad library is public, so is Google's, and anyone can go and look. What you can't get for free is the rest of it: three platforms in one place, a watchlist that tells you when a competitor launches something new, run duration and reuse count attached to every headline, the creative file as a download, and the whole thing reachable from an API. By hand it means opening each library separately, scrolling, and still not knowing how long anything has been running.
What AdKit Actually Does
It's worth laying out the pieces, because "AI ads tool" could mean almost anything.
The ad library. When I queried it with no filters it returned 833,134 ads, of which 294,510 were on Meta. The same query had returned 824,533 the day before, so it's growing by several thousand a day. It's built on the public ad libraries from Meta, Google and LinkedIn. You add competitors to a watchlist and their ads accumulate.
MCP access. AdKit has no chat window of its own. It connects through MCP, so Claude or ChatGPT calls its functions directly. In practice you type "target the US, put it on Instagram" in whatever AI you already use, and the campaign gets built. This also means performance reporting comes back into your chat. I never downloaded a report or screenshotted a dashboard to feed my AI. I just asked how yesterday went.
The AI ads generator. AdKit calls it Studio. It writes copy and generates static image ads.
Campaign management. Create, edit, publish and pull results across Meta, Google, TikTok, LinkedIn and Reddit. The three nested layers of campaign, ad set and ad all get built through conversation.
Activity alerts and swipe file. A dated timeline of new ads your tracked competitors launch, and a place to save what you want to steal from. This is the screen I'd open daily.
Ad Longevity Is the Metric Nobody Talks About
The most useful thing in the library turned out to be the dates.
One competitor, Stardust, had a single video that had been live for ten months straight. I can't see their revenue. But nobody keeps paying to run the same video for ten months unless it's working.
That reframed how I read the whole database. I stopped browsing everything and started filtering: ignore every ad except the ones that have been live six months or longer. A headline still running after eleven months is a sentence that company has paid to validate. Ones that ran briefly and vanished probably failed. You don't get performance metrics directly, but "survived a long time" is a good proxy for "works." AdKit's own Ad Score runs on the same logic, weighting run duration and reuse frequency.
So my process now is three steps. Put every competitor I can name on the watchlist, and if you run out of names, AdKit suggests more. Filter to the long runners. Write my own version of that angle. I'm not guessing what to say anymore; I'm reading what already survived.
It Makes the Creative Too
Two ways, and they're at different maturity levels.
Static images work well. Studio generated three PairCare ads I could have run without a redesign, built around a "Birth control isn't a solo job" headline. One thing to know: it runs on your own OpenAI key, so images are not included in the subscription. In my case it came to about fifteen cents each, and the cost is displayed.
Each creative moves through Plan, Create and Preview. The Plan tab is effectively a creative brief: concept, target audience, format, aspect ratio, destination platforms and brand colors. You can pull reference material straight from a competitor's ad in the library. It's genuinely useful as a spec sheet to hand a contractor.
Video generation isn't there yet. You can select Video as a format, but the generation screen says "Coming Soon."
So for video I used the library differently. I downloaded six competitor ads, including that Stardust video, and handed them to my own creative agent with instructions to keep the structure and pacing and change the angle. The video itself was generated with Kling AI. That combination, AdKit for reference and something else for generation, worked fine.
The Four Day Experiment I Ran
I set two things before connecting anything, and I'd tell anyone to do the same.
First, agent permissions. AdKit gives you four levels, from read only up to letting the agent publish campaigns on its own. I like that it's a choice, since different people want different amounts of autonomy, but it arrives on the most autonomous level by default, so decide before you connect. It's in Settings, then Agent Permissions, then Change access. Thirty seconds.
Second, a spending cap on the ad platform itself rather than inside AdKit. I set mine to about eighteen dollars. That became the absolute maximum I could lose regardless of what the agent did.
Then I ran it: US only, Instagram only, about five dollars a day, traffic objective pointing at paircare.love. Two creatives, one built around the pill reminder angle and one around the new cycle tracking feature, because I wanted to know which one people actually came for.
Where It Broke
Three things, and two of them weren't AdKit's fault.
TikTok wouldn't let me target the US. My ad account is registered in Korea, and TikTok blocks Korean accounts from targeting US users. This is a restriction on the account, not the ad, so there was nothing the agent could do. AdKit surfaced TikTok's raw error verbatim with a retry button, which is exactly what I want from an error, but the check happens at publish time. Knowing beforehand would have saved me two days.
"Active" doesn't mean "delivering." My first Meta ad reported publish success and showed active, but spend stayed at zero. Meta had rejected it, because I'd pointed a traffic campaign at an App Store URL and that requires the app install objective, and the rejection never reached the agent. My setup error, but I lost time because the failure was silent. Open the platform dashboard yourself after the first launch of any campaign.
TikTok budgets display in dollars on a won account. ₩50,000 showed as $50,000.00/day. Meta accounts display correctly on the same screen, so it's a bug in the TikTok path rather than a global hardcode. Cross check budget numbers in TikTok Ads Manager.
What $14 Actually Bought
Four days, $14.15, then I shut it off.
That bought 1,283 impressions, 162 clicks, and 127 landing page views. A click through rate of 12.6%.
Free trials from all of it: zero. Not one. This was my first time running paid ads, so I genuinely don't know whether that's a normal conversion rate for a cold audience at this volume. What I can say is that fourteen dollars proves nothing about revenue.
Then there's the part I got wrong. I put both creatives in the same ad set, thinking that was an A/B test. It isn't. By the end of day one Meta had shown the pill reminder ad 349 times and the cycle ad 20 times, and it never went back. Four days later the second ad had 47 impressions in total.
Meta wasn't picking at random. More people watched the first seconds of the pill video, a 67% hook rate against 55%. That's a real signal. But look at what it had on the other side: twenty impressions and one click. Run the statistics on one click out of twenty and the true rate could be anywhere from zero to about twenty percent, which comfortably contains the winner's 12.9%. The data never showed the second ad was worse. It showed that nobody knew yet, and the moment Meta picked, it stopped collecting the evidence that could have changed its mind.
If you want a real comparison: one creative per ad set, with separate budgets.
Meta Has Its Own Free Ads MCP Now. Is AdKit Still Worth $29?
This is the question I'd want answered if I were reading this, so I'll take it head on.
Meta launched its own Ads AI Connectors in open beta on 29 April 2026: an official MCP server, twenty nine tools, free for the duration of the beta. If all you want is to drive your own ad account from a chat window, you can do that today without paying anyone.
So that part of AdKit is no longer scarce, and it would be dishonest to pretend otherwise.
There's also the question of what sits between the agent and your live account. With Meta's connector the agent is working on Facebook itself, so a mistake lands on the account you run your business from. AdKit keeps a draft step in the middle, where the agent proposes and you approve. Meta's connector is also narrower in what it covers, and it's still a beta that not everyone has.
What Meta's connector cannot do is the other half. It talks to your account. It won't tell you that a competitor has been running the same video for ten months, or hand you the mp4, or show you the headline that's been live since last year. That's a different dataset, assembled from the public ad libraries of Meta, Google and LinkedIn, and it's the part that actually changed what I did.
So the honest framing: you're not paying $29 for the MCP. You're paying for the library, and the MCP is how you use it without leaving your chat window.
Pricing: What Basic Covers
I'm on Basic, $29/month or $228/year, which works out to $19/month if you commit annually. Its features are scoped to one project, meaning a single brand workspace and its ad accounts: the ad library, competitor tracking, activity alerts, the AI ads generator, the swipe file, MCP access, and 25 ad analyses per day.
The tier above is $49/month or $408/year and adds unlimited projects. Same features otherwise. A $20/month gap is small enough that if you run several apps it's barely a decision.
There's a seven day free trial with no card required until it ends. Spend that week inside the competitor library before committing to anything, because that's the part that would decide it.
Two things aren't included. Image generation bills to your own OpenAI key, and ad spend is obviously separate. My whole experiment came to roughly $43: the subscription, forty five cents of images, and $14 of actual advertising. Most of what I paid was the tool, not the ads.
Final Thoughts
Ahrefs isn't valuable because it writes your articles. It's valuable because it shows you what's already working for everyone else. AdKit sits in that same seat for advertising, except it doesn't stop at the data. It can execute too.
That's the honest frame. There's no brain inside it telling you which angle to pick. It hands you the raw material and the controls, and the strategy is still yours, or your LLM's.
Would I keep paying for it if nobody were paying me to write this? That depends on one thing I haven't decided. If I commit to running paid ads properly, then yes, without thinking about it. I'd be in the competitor library every week. And it isn't only the library you're paying for, it's the safety layer and more features on top. It earns the $29 on its own. If I decide to grow organically on TikTok first, it's the wrong tool, and not only because the subscription would sit unused. The library doesn't index TikTok at all, and TikTok's own public library only covers Europe and the UK, so there is no version of this that tells you what your competitors are running there.