Google quietly uploaded an app called COSMO to the Play Store on May 1. Within hours, it pulled it. But not before developers downloaded the 1.13 GB package, tore it apart, and found something that makes Google Assistant look like a parlour trick: a fully on-device AI agent powered by Gemini Nano that listens to your real-world conversations and takes action before you even ask.
This wasn’t a chatbot. This wasn’t another “Hey Google” trigger. COSMO is what happens when Google stops waiting for you to talk to your phone — and starts letting your phone talk to itself about you.
What COSMO Actually Does — And Why It’s Terrifying
The app’s internal documentation, unearthed by Android developers within hours of the listing going live, describes a system built around “proactive skills.” These aren’t voice commands. They’re triggers that fire based on what COSMO overhears in ambient conversation.
Two examples found in the code tell the story. A Calendar Event Suggester detects when you’re verbally agreeing on a meeting time with someone nearby — and offers to schedule it. A Quick Photo Lookup notices when you mention a specific photo in conversation and surfaces it from your gallery without being asked. These aren’t hypotheticals. They’re built, labelled, and functional inside the app package.
The architecture runs on three fulfillment modes: Hybrid (cloud when online, Nano when offline), PI Only (server-side processing), and Nano Only (fully local). The fact that Google built a complete offline path means this isn’t just a cloud experiment with a local fallback. It’s designed to work when you have zero connectivity — which means it’s designed to listen everywhere.
The 1.13 GB Download That Google Didn’t Want You to See
COSMO’s file size — 1.13 GB — is the tell. That’s the weight of a bundled Gemini Nano model sitting entirely on your device. No server round-trips for basic inference. No latency. No data leaving your phone for the core listening loop. Google essentially shipped a local brain that processes ambient audio in real time, extracts intent, and maps it to actions.
The privacy implications cut both ways. On one hand, on-device processing means your conversations aren’t being streamed to Google’s servers. On the other, a phone that actively processes everything said around it — even locally — is a fundamentally different product than one that waits for a wake word. Google knows this. That’s almost certainly why they pulled it.
This Is Google’s Answer to Apple Intelligence — And It Goes Further
Apple has spent the last year building Apple Intelligence as an on-device AI layer that processes personal context. But Apple’s approach is reactive — you invoke Siri, you ask a question, you get an answer. COSMO flips that model. It doesn’t wait. It listens, infers, and offers. That’s a philosophical gap, not just a technical one.
Apple bets that users want AI that responds on command. Google is betting that users want AI that anticipates commands. The question is whether users are ready for a phone that effectively eavesdrops — even if it does so locally and never phones home.
The timing here matters. Google I/O 2026 is weeks away. COSMO’s basic interface and internal labelling suggest it’s a developer testing tool, not a consumer product. But the architecture is consumer-grade. The proactive skills are consumer-use-cases. And the fact that Google built a full offline mode suggests they’re not testing whether this works — they’re testing whether people will accept it.
Why Google Pulled It — And What Happens Next
Google yanked COSMO’s Play Store listing within hours. Everyone who didn’t grab it in time now gets a “not found” message. The company hasn’t commented publicly, which is standard Google behaviour when an internal tool escapes into the wild ahead of schedule.
But the damage — or the preview, depending on your perspective — is done. The developer community has the APK. The internal documentation is public. And the feature set is clear: Google is building an AI assistant that doesn’t need to be asked. It just knows.
The likely scenario is a polished version of COSMO debuts at I/O as part of a broader Gemini on-device push. Google has been telegraphing this move for months — Gemini Nano integration in Pixel phones, the expansion of on-device AI processing, the push toward agentic capabilities. COSMO is the product those investments were always building toward.
The Real Question Nobody’s Asking
Every tech company is racing to build AI agents that act on your behalf. But there’s a difference between an agent you summon and an agent that’s always on. COSMO is the latter. It sits in your pocket, processes ambient audio through a local LLM, and decides — on its own — when to surface information or suggest an action.
That’s not an assistant. That’s a coworker who’s always listening to your side of every conversation and occasionally tapping you on the shoulder. Some people will find that indispensable. Others will find it deeply unsettling. Google clearly hasn’t decided which group is bigger — otherwise they wouldn’t have pulled it.
The verdict: COSMO is the most important AI product Google has accidentally revealed in years. Not because of what it does today — it’s clearly unfinished — but because of what it says about where Google thinks the AI assistant category is headed. The wake-word era is ending. The always-listening era is about to begin. And Google just showed us the blueprint, whether it meant to or not.
