The dream is simple: you're away from your desk, you remember the bug, and you want the fix underway before you're back. The usual answer — SSH from a phone into a tmux session and drive a terminal with your thumbs — is technically possible and practically miserable.
The terminal was never the point. The task was. So skip the terminal.
The phone-shaped version of coding agents
What you actually need from a phone is four things, none of which is a shell:
- Pick a task that's already well-defined.
- Fire it off to a machine that isn't your laptop.
- Get pinged if the agent needs a decision.
- Review the result when you're somewhere with a real screen.
That's exactly the loop Cloud Sandboxes implements — and every step of it works from a browser on your phone:
- Your project's scope is already there. Tasks with acceptance criteria, defined when you had a keyboard and your full brain. From your phone you're not writing a prompt — you're tapping run on a task that's ready.
- The sandbox does the heavy lifting. A dedicated VM spins up with your repo, your plan, and your connected agent — Claude Code or Codex, on your own account, no markup.
- The run streams back live. Status, logs, and the agent's comments land in the Command Center. Glance at it like you'd glance at a delivery tracker.
- Questions come to you. If the agent hits an ambiguity — "should this endpoint be paginated?" — it flags the task and pauses. You reply in the task's comments, from the couch, and the run resumes. No session to keep alive, no connection to babysit.
- It ends in a pull request. Work lands on its own branch. You review the diff on GitHub whenever you're ready — nothing merges without you.
What this replaces
| The old way | The phone way |
|---|---|
| SSH client + tmux + thumb-typing | Tap a task, tap run |
| Terminal scrollback on a 6-inch screen | Live status + comments feed |
| Agent silently guesses when ambiguous | Agent pauses and asks; you reply async |
git diff over SSH |
A normal PR review, later |
The honest catch
This only works because the task was scoped before you left your desk. A phone is a terrible place to write a spec and a great place to pull a trigger. The setup — scope the project, connect GitHub, connect your agent CLI in Settings → Cloud Agents — happens once, with a keyboard. After that, dispatching real work from a parking lot is genuinely a ten-second operation.
That's the whole trick: move the thinking to when you're sharp, and the dispatching to wherever you are.
Setup guide: how cloud runs work end to end → · Included on every paid plan, from $29/mo.