Google accidentally leaks COSMO on the Play Store

Most mobile AI assistants still sit quietly waiting for you to tap an icon, but Google's researchers are already testing what it looks like when your phone starts running errands for you.

As reported by 9to5Google, Google accidentally published an experimental Android app called COSMO to the Play Store. The package name traced back to Google Research, but because it slipped out under Google's main developer account, the internet spotted it before it was swiftly pulled.

Packing Gemini Nano on-device

COSMO weighed in at a hefty 1.13 GB download. That bulk comes from bundling a local Gemini Nano model directly inside the app.

Once installed, it asks for several broad system permissions, including screen access and voice matching. The user interface is barebones chat, clearly built as an internal playground for Google researchers rather than a polished consumer release.

Inside the settings menu, users could pick between three fulfillment models. There is a local-only Nano mode, a server-only PI mode, and a hybrid setup that leans on the cloud when online and falls back to Nano offline.

From a chatbot to an active agent

What makes COSMO fascinating is its list of experimental skills. Google is clearly shifting from passive question-answering to proactive phone automation.

For heavy lifting, the app features Deep Research to pull together multi-source reports, alongside a Browser Agent that uses a tool named Mariner to automate web tasks. When you change context, a conversation summary feature wraps up what you just talked about.

Then there are the subtle context-aware skills. If you mention plans with a friend, it offers to schedule the calendar event. If you mention a photo you took earlier, it can surface it without interrupting your chat, while a Recall feature helps dig up things you are trying to remember.

Why it matters

We think this leak offers an unfiltered look at where Android is ultimately headed: transforming the entire mobile operating system into an omnipresent personal agent.

The hybrid architecture makes a ton of sense. Running Gemini Nano locally handles low-latency context awareness, screen reading, and offline tasks, while passing massive research workloads up to server-side models.

That said, the hurdles are obvious. A 1.13 GB footprint and background model execution will chew through battery and memory on midrange hardware. More importantly, granting an experimental agent deep access to your screen and conversations will trigger major privacy questions.

You cannot download COSMO today, but you do not need to. It proves that Google is already building the layer where your phone stops waiting for commands and starts anticipating your next move.