I'm 16 and built a desktop app that controls your PC autonomously — you type a task, it executes everything.

Most automation tools still need you to set up workflows, connect APIs, or write code.

I wanted something different — you just describe what you want in natural language, and it does it.

So I built Flex. It takes full control of your mouse, keyboard, and apps to execute complete workflows autonomously. Works on Windows and Mac.

Real example:
"Check the latest email about March revenue, update the Google Sheet, and post the chart to Discord."

It opened Gmail, read the email, updated the spreadsheet, generated the chart, and posted it to Discord — in 90 seconds, zero clicks from me.

Watch the 90-second demo: https://www.loom.com/share/04c27c6a5e1d48a29e38d8b2c85c7622

Private beta is open — secure your spot now:
👉 https://flexagent.app

Under the hood — the part I'm most proud of:

The obvious approach (Claude Computer Use, etc.) relies on continuous screenshot-based vision. It's slow, expensive, and error-prone with coordinates.

I built a 3-Tier Execution Engine that makes vision a last resort:

Tier 1 — OS Accessibility Tree: Direct hook into the OS UI tree. Zero-latency, zero vision tokens. This handles most tasks instantly.

Tier 2 — Set-of-Mark (SoM): If the accessibility tree is blocked (custom canvases, games, etc.), I take a screenshot and number every UI element. The model picks a number — no coordinate guessing, no misclicks.

Tier 3 — Raw Vision: Only if both fail, the model calculates raw X,Y coordinates itself. This is what most agents start with. I use it as a last resort.

On top of that:

• Mouse moves on Bezier curves with variable speeds (human-like, bypasses bot detection)
• Typing is character-by-character
• 50+ built-in tools: terminal execution, file management, isolated browser
• Ultimate Mode: complete human interface takeover across any app

I'm 16, built this entirely solo. No team, no funding, no co-founder.

Happy to answer questions about the build.

https://flexagent.app