Google's leaked COSMO app is the first AI I might actually trust with my files

by · Android Police

I’m officially burnt out on chatbots. Typing prompts into a text box feels like learning a programming language to do the work my phone should already understand.

That’s why Google’s leaked COSMO app is the first mobile AI idea in a while that has my attention.

The leak points to an assistant built around screen context, with enough awareness of what you’re doing to suggest the next step.

Google accidentally pushed COSMO live through its official Play Store account on May 1, 2026, then pulled the listing within hours.

With Google I/O getting close, I can’t help but wonder if COSMO was headed for the keynote. I’m definitely curious to see what Google says next.

Related

Chrome AI takes up 4GB and there may be nothing you can do about it

Delete it and it could come back

Posts 2
By  Chandra Steele

Google’s leaked assistant may not need the cloud for everything

COSMO’s size is the first big hint. The app reportedly needs a 1.13GB download, which is unusually large for a regular assistant app.

That probably comes down to Gemini Nano. The software appears to include a local version of Google’s small AI model, so some AI work can happen on the phone instead of going to Google’s servers.

Local AI isn’t free, though. Running a model directly on the phone can drain the battery if it isn’t managed well.

COSMO’s leaked settings include a section called Fulfillment Models with three options.

  • Nano Only runs locally for privacy and offline use.
  • PI Only relies on a remote Google server for heavier reasoning.
  • Hybrid appears to be the default, switching between local and cloud processing depending on the task and processing requirements.

Screen awareness could be Google’s mobile AI advantage

Right now, using AI on a phone often means stopping what you’re doing, switching apps, and spelling out context to a language model.

COSMO’s bet is that the phone already has the context, and it only needs permission to see it. The app requests Android’s AccessibilityService API.

Google designed that system-level hook for tools such as screen readers, but here it likely gives the mobile agent a way to see what’s on the display.

If your messages, browser tab, or calendar are visible, COSMO could use that on-screen context to understand what you’re doing.

A reported example is scheduling. If you’re texting a friend about dinner, COSMO can recognize the intent and surface a prompt to add the event to your calendar.

The permission side is still a little scary. Accessibility access can expose a lot of sensitive on-screen content.

In a cloud-only assistant, that would feel too invasive for many people. COSMO’s local Gemini Nano model is what makes this leak feel less creepy than it otherwise would.

If it can analyze screen context on the device, at least some sensitive data can stay away from Google’s servers.

The boring skills are what make COSMO interesting

Google has built COSMO around 14 skills:

  • List Tracker: Suggests a list when it notices you are planning something.
  • Document Writer: Offers to write a draft when you mention needing a document.
  • Calendar Event Suggester: Notices a plan with a time attached and offers to put it on your calendar.
  • Browser Agent: Uses Mariner to complete tasks in the browser.
  • Add Timer: Spots a task with a time limit and suggests a timer.
  • Deep Research: Offers a deeper report when a simple answer is not enough.
  • Quick Photo Lookup: Finds the photo you are talking about so you can share it faster.
  • Google it: Suggests searching the web when that would answer the question.
  • Jargon Definitions: Explains unfamiliar technical terms as they appear.
  • Provide Insight: Adds helpful background while you are reading or discussing something.
  • People Understanding: Gives you context about a person.
  • Event Understanding: Gives you context about an event.
  • Recall: Finds something you seem to be trying to remember.
  • Conversation Summary: Summarizes a recent conversation when you switch to something else.

These are ordinary chores, which is why they’re promising. A lot of phone productivity is still based on small loops of manual work. COSMO’s skills are built around collapsing those.

Probably a test bed and not a new Gemini app

Now for the cold water. Everything about the leaked build points to something unfinished.

Subscribe for deeper newsletter analysis of COSMO

Dig into mobile AI by subscribing to the newsletter for focused coverage and analysis of developments like COSMO, local on-device models, and screen-aware assistants — context that clarifies how these features reshape phone productivity.

Get Updates

By subscribing, you agree to receive newsletter and marketing emails, and accept our Terms of Use and Privacy Policy. You can unsubscribe anytime.

The chat interface is bare bones.

The screenshots attached to the accidental Play Store listing were squished into the wrong aspect ratios.

Nobody should consider this a polished consumer product yet, and I doubt Google ships COSMO as a standalone Gemini replacement.

My reading is that Google probably uses COSMO as a test bed for the skills and then folds the parts that work into Gemini or Android over time.