LocalPilot

Why I built it

I wanted to explore how far I could take AI coding tools with a complex project. The idea was to give a locally running model access to a Linux computer that I could also use from my phone.

What it does

LocalPilot has two views: Agent, where I can give the model a task and follow its progress, and Desktop, where I can watch the Ubuntu screen and take control myself.

The model can work with files, run approved terminal commands, and use Firefox. I can review actions, cancel tasks, or pause the AI while I use the computer. Task history, file backups, and connection recovery make it easier to pick things up again.

How it connects

The iPhone web app connects privately through Tailscale. I can assign tasks, review actions, and see the shared desktop.

My role

I chose the approach and model, directed revisions, checked outputs, troubleshot problems, and tested the system from my phone. AI tools wrote most of the code. For me, the project was an experiment in directing those tools and understanding whether the resulting system actually worked.

What I tested

I tested remote access over cellular data, manual desktop takeover, Firefox browsing, software installation, and launching the app from my iPhone Home Screen. The project also includes 109 passing automated tests and six selected practical task evaluations.

This is a personal prototype. Those checks cover the tested tasks, rather than general reliability on long or unfamiliar work. The laptop needs to be awake and online, and the system still needs supervision.