Driving Claude Code from my phone: see it, point at it, say what's wrong

I run Claude Code on my Mac and drive it from my phone. The agent does the work. My job is the part it can’t do for itself: looking at what it made and deciding what’s wrong with it.

On a laptop that part is free. The browser is right there next to the terminal, so you glance over and you know. On a phone it has always been the expensive part. The agent rebuilds a page, and to actually see it you open a tunnel to a browser, or you deploy the branch just to look, or you wait until you’re back at the desk. You end up directing a frontend you can’t see.

So the loop I care about is the one where I can see it. It goes: see, point, say, see.

See

You run your dev server in the terminal like you always do. npm run dev, bun dev, whatever you use. Then you open the preview, and it slides up over the terminal at half height. The top half is your live Claude Code session. The bottom half is your actual site, loaded from the dev server on your machine.

You did not start a tunnel. You did not deploy. The page is your localhost, reached over the SSH connection the terminal already holds. When the agent edits a file, the dev server’s hot reload pushes the change down the same connection and the preview repaints. You watch the page change a second or so after you tell the agent to change it.

Swipe up and the preview fills the screen in a clean mobile view. Swipe down and you’re back to both halves. The terminal stays live the whole time, so you’re never trading the session for the render.

Point

Two things are worth pointing at. The page, and everything else the agent makes.

The page is the preview you already have open. Everything else is files. When Claude Code writes or edits a file, it prints the path in the terminal:

⏺ Update(src/components/Hero.astro)
⏺ Write(docs/launch-plan.md)
# tap either path and it renders in a sheet

Tap one of those paths and the file renders. A markdown plan comes up formatted, not as raw text. A logo it generated comes up as the image. A standalone HTML mockup renders as the page. A PDF opens as the PDF. It reads the file straight off your machine over the connection, so this works whether or not a dev server is running. You only ever tap what the agent printed, so it stays a view of the work, not a file browser.

Say

This is the step that used to be words. Something on the page is wrong, the spacing is off, a button sits in the wrong place, and you have to turn that into a sentence the agent can act on.

Instead, you snapshot the preview and circle the thing with your finger. The circle goes down in orange. You add a few words, “this gap is too big,” and send. It lands in the agent’s prompt as the screenshot, marked up, plus the file path. The agent sees the image you drew on.

See it change

Then the agent reads the marked-up screenshot, changes the file, and the dev server repaints the preview. The gap closes, on the screen, in front of you. If it’s right, you move on. If it’s not, you circle it again.

That’s the loop. See the render, point at the problem, say the one thing that’s wrong, see it change. All of it on the phone, over one SSH connection, with nothing installed on your machine but the SSH server it already has.

I built this because it’s the loop I wanted for my own work, and I run it every day. The agent writes the code. I’m still the one who has to look at the result and decide, and now I can do that from wherever I am instead of saving it for the desk.