I gave Gemini 4 a Mac, a phone, and $1,000 to run a software company (Day 1)

I gave Gemini 4 a Mac, a phone, and $1,000. For seven days it runs a software company and ships one open-source app.

Gemini-4-Argon came out today. I tried it on long-horizon planning, and on handing work to other tools. A chat box is the wrong job for it.

I'm doing this in public for seven days. I gave Gemini 4 a Mac, its own browser, a physical phone, and a crew running on DeepSeek. It has to run a software company. By day 7 it ships one thing, built from scratch: a free, open-source alternative to a known paid app. It picks the app.

The Mac and the phone on the desk

$1,000, split in two

Multi-agent setups waste money in two familiar ways. Put a frontier model on every click and the budget is gone in a couple of days. Put a cheap model in charge of the plan and the plan wanders.

This one runs on Munk AI, a local runtime I wrote. The token cap is $1,000.

$800 is Gemini 4. It keeps the plan for the whole company, writes the core, decides what done means, and messages me for a yes or a no.

$200 is DeepSeek. It does the repetitive work.

Testing a screen takes dozens of passes: look, click, see what changed. That eats context. The runtime finds the controls and the OCR coordinates first, and DeepSeek drives the loop from there. Bugs go back to Gemini 4.

That same crew makes the images, cuts the promo videos with FFmpeg, and posts and replies from the browser. Some of those posts go to Xiaohongshu.

More than code

If it only writes code, it is an IDE plugin, not a company.

  1. The Mac. If an app has no API, it looks at the screen and clicks the button anyway.
  2. Its own browser, and a phone on the desk. It uses the browser to look up competing products and docs, and to post. The phone is where it tests apps. Later the phone can read its own SMS codes.
  3. No merge without a test. If the build fails, or a real click-through on the device fails, the change does not merge.

What I do

For seven days I write no code and publish nothing myself.

I tap approve or reject on my phone. A first login or a payment comes to me. I type the password, or I pay. The password never goes into the prompt. Then it carries on.

On day 7 the alternative is open-sourced, finished or not. The setup is published with it, so someone else can run it the same way.

Day 1

Today is 14 October 2026. Day 1 of 7.

I told it to choose the product. The crew uses the browser to find paid software that people already dislike, and that has no obvious open-source alternative. Tonight I get three proposals on my phone.

If you pay for an app like that, tell me which one. The names that come up often go into the same review. On day 2 I name the app, and the code starts.