Claude Code on your iPhone, running on your server.

The agent doesn't stop when your screen does.

An hour-long refactor shouldn't depend on an app staying in the foreground, and here it doesn't. Run Claude Code or Codex inside tmux or herdr on the server; when iOS suspends dotMux, the session carries on without you. Open the app again and it re-attaches to the same windows, so you read what the agent did while you were away.

Your prompt history, the agent's context and every file it touched stay on the server. The phone only draws the terminal.

Show it the bug, don't describe it.

A headless server has no clipboard for an agent to read, so dotMux bridges one. Copy a screenshot on the phone and paste: the image goes up over SFTP on the connection you already have, and its file path is typed into the prompt. Claude Code turns that path into an attached image. Other CLIs get the path too, to use with their own file syntax.

Location data is stripped before upload, and old uploads clean themselves up on the host.

Take its output with you.

Claude Code wraps its transcript to fit the screen, so copying the screen as shown would break every long line. Long-press and drag over it instead: Claude Code highlights the text itself, and the copy lands on your clipboard as logical lines, ready to paste into an issue or a chat. That handoff needs tmux 3.4 or newer on the host; under herdr, Copy Joined rejoins the lines instead.

For long replies, the composer drafts a multi-paragraph prompt and sends it in one paste. Esc and Ctrl sit on the helper bar for interrupting a run that went the wrong way.

One agent per machine, all of them live.

Keep a build box, a cloud VM and a homelab connected together. Switching between them doesn't reconnect or ask for Face ID again: the pill in the tab bar lists every open host, and a tap puts you in front of the agent on the next one.