Building Apps Is Solved. Distribution Isn't.

Full analysis

Mike Khristo, who runs the Aperiam-backed startup Layers, went on Corey Ferengul and Joe Zawadzki's podcast to argue that AI has made building an app trivial and getting it downloaded the only hard part left. His fix: a new "Growth MCP," a plug that lets an AI coding assistant like Claude or Cursor reach out and run marketing tasks (research competitors, post to TikTok, launch campaigns) without opening a separate website.

Nothing here forces a decision for a large ad-tech operator. This is a founder pitch. But it carries one real signal worth pulling apart: when Khristo's team wired their tools into an AI agent, the agent didn't want the "do the thing" buttons. It wanted the data.

What's actually being decided: not "should I buy Layers." The question underneath is whether the tools you already pay for (analytics, attribution, campaign managers) are about to get bypassed because your team starts working inside an AI assistant instead of logging into dashboards. That's easy to undo if you're wrong, and cheap to test. No deadline sets the clock here. Nobody is turning anything off.

The Skeptic. Almost every number in this episode comes from the company selling the thing. "13 trillion seconds of TikTok video, 23 million posts, posting once or twelve times a day doesn't matter." Nobody outside Layers has seen that data. And the convenient part is that the finding sells the exact feature Layers ships: automated high-volume posting. Khristo's swipe at Meta, that Facebook's ad manager keeps its AI creative mediocre on purpose to dodge blame, is his word "suspicion," and he says so. The plainer reason is that generic AI creative is just hard to do well. Take the whole episode as a smart pitch, weighted accordingly.

The Builder. Strip out the sales and one thing is genuinely useful: an MCP is just a standard way for your AI assistant to call an outside tool. If your marketers live in Claude or Cursor all day, giving the agent a live feed of "here's what's working for competitors right now" beats making it click through a dashboard. The finding that agents reach for research over execution is real and matches what I see on Tuesday mornings. The agent can already write a campaign. What it lacks is current market data. That's a hint about where any tool you buy should put its value: fresh data the model can't get on its own, not buttons the model already knows how to press.

The Open-Source Advocate. MCP is the good news buried in the pitch. It's an open standard, not Layers' invention. Anthropic published it, and anyone can build a plug for it. So the useful takeaway isn't "buy Layers' MCP." It's that the tools you already own should expose one, and if they won't, a competitor's will, or you wire one yourself. The moat Khristo claims is his proprietary TikTok dataset. Fine. But the connection layer is free and standard. Don't pay a premium for the pipe when only the water is scarce.

The Compute Pragmatist. One offhand item matters more than the marketing tactics: Khristo points at FAL's live product that generates video faster than it plays back. If that holds up, the cost of making organic social video drops toward zero, and the constraint on cranking out TikToks stops being production and becomes distribution and quality. That's the actual engine behind his whole thesis. It also means the flood of AI video gets worse for everyone relying on organic reach, because when everyone can generate infinite clips, the platforms will tighten what they distribute. His "posting frequency doesn't matter" finding is measured in today's world. It won't survive a world where video is free to make.

Where they disagree. The Builder sees a real workflow shift: work moves into the agent, dashboards fade. The Skeptic says fine, but that shift helps whoever owns the data feed, and Layers is a tiny player claiming a moat it hasn't proven. The Open-Source Advocate splits the difference: the workflow shift is real and it's built on a free standard, so the incumbents Layers is trying to displace, AppsFlyer and Branch, are better positioned to win it than any startup that happened to ship an MCP connector first. That points at the incumbents, not at Layers.

What this hinges on. One belief: does real work actually migrate from dashboards into AI assistants over the next year, enough that a tool without an MCP loses users? If yes, every analytics and attribution vendor you use needs to expose one, and the ones sitting on the most data win. If no, this is a niche convenience for indie app developers and nothing more. The cheap test costs you an afternoon: wire one MCP-enabled tool into your team's Claude or Cursor setup, watch whether anyone actually uses it a week later. If they drift back to the dashboard, the thesis is early.

The council leans toward "real trend, wrong protagonist." The move to agent-first workflows is happening. Layers being the one to win it is not the bet.

Prediction: By the time AppsFlyer or Branch ships its next major platform update before mid-2027, at least one of the two established mobile attribution vendors will publicly release an MCP connector letting AI agents pull their measurement data directly, neutralizing the "first mobile measurement agent" gap Layers is claiming.

Confidence: Medium. MCP is free and standard; incumbents own the data that makes it useful.

Why: Khristo's own finding is that agents want live data, not execution buttons, and the data is exactly what AppsFlyer and Branch already have and Layers is scrambling to rebuild. MCP is an open standard from Anthropic, so exposing one is a small engineering job, not a research project, which means the barrier for an incumbent is a quarter of work, not a moat. Once agent-first workflows show up in customer requests, a vendor sitting on years of attribution data has every reason to add the one connector that keeps it visible inside Claude and Cursor, and no reason to cede that surface to a startup. The opposite outcome, both incumbents ignoring MCP entirely, would require them to watch their tools get bypassed and do nothing, which is not how vendors defending a data business behave.

Revisit by 2027-06-30: We're right if AppsFlyer or Branch ships a documented MCP connector for its attribution data by then. We're wrong if neither does and agent access to their data still runs only through legacy APIs or the web dashboard.