Three ways to code from your phone, and what each costs

A phone is a poor place to run a coding agent and a fine place to steer one. Every setup that works puts the agent on a computer that stays on and makes the phone a window onto it. Three versions come up again and again: a computer at home reached over Tailscale, a cloud VM that's always there, with mosh and tmux, and a VM billed by the hour that sleeps when nobody uses it.

They differ in what you keep running, what happens when the train goes into a tunnel, how you learn the agent is waiting, and what you pay. If you've picked one and want the Claude app as your window, our Remote Control post covers the setup.

What every setup needs

• A computer that stays on. The agent's commands run there. If it sleeps, the agent stops.
• A session that outlives the connection. Anthropic's Remote Control docs give the rule for any remote terminal: start claude inside tmux or screen, so it keeps running after SSH disconnects. The tmux manual says its sessions survive an accidental disconnection, such as an ssh timeout.
• A way in: an SSH app, a browser or the Claude app.
• A way to hear from the agent, so you don't learn an hour later that it stopped to ask something.

1. A computer at home, over Tailscale

Ryan Bergamini's doom-coding guide is the simplest version. A computer at home has SSH on and sleep off. Tailscale runs on it and on the phone, and Termius on the phone connects over SSH, where you run claude. His list of requirements: a computer running around the clock, a smartphone and a Claude Pro plan.

Tailscale puts the phone and the computer on one private network, and its docs say a device keeps its Tailscale address when it moves from Wi-Fi to a cellular network.

The weak point is the computer. Bergamini's troubleshooting section lists how his stopped answering: it had to be unlocked after the power came back, and a closed laptop plugged into a monitor dropped off Tailscale after about ten minutes, with sleep turned off. His guide also runs claude straight in the SSH session. Add tmux, and a dropped connection costs you a reconnect instead of the agent.

2. An always-on VM with mosh and tmux

Granda's Claude Code On-the-Go moves the computer to Vultr: an 8 vCPU, 32 GB VM with no SSH on its public address, reached only over Tailscale. Termius connects with mosh, the shell attaches to tmux on login, and six agents work in parallel on git worktrees.

mosh is the part built for phones. mosh.org presents it as a replacement for interactive SSH that keeps you logged in as you change networks, and while the client sleeps. When the network drops, it warns you and waits for it to come back. As Granda puts it, "Mosh is the key—it survives network transitions." Its documented limits:

• It shows only the visible screen; mosh.org's fix for scrollback is tmux on the server.
• It needs UDP ports 60000 to 61000 open on the server.
• It doesn't forward your SSH agent, so Granda runs git commands that need GitHub over plain SSH.

He prices the VM at $0.29 an hour, about $7 a day while it runs, and halts it when he stops work. If you copy that, check how your provider bills a stopped machine: Vultr's billing docs say a stopped instance keeps billing until you destroy it. On Hacker News he agreed with a commenter's estimate of $210 a month and said a smaller instance would also work.

3. An hourly VM that sleeps when idle

The third setup keeps the machine in the cloud and stops paying for idle hours. shedix works this way. Its machines come with Claude Code, Codex and tmux installed (all but Plain ones, which are the bare OS), and you reach one from a phone through its web terminal, through Remote Control in the Claude app, or over SSH. The terminal and agent tabs on the machine page run in tmux on the machine.

A machine sleeps after 60 minutes unused: no SSH session or open terminal tab, no agent writing its conversation, and its CPU and network quiet. Sleep is a stop. The disk is kept, running programs end, and there's no hourly charge until you start it again, which takes under a minute.

An agent waiting for your answer doesn't count as use. If nothing else does, a question left for an hour puts the machine to sleep, and Remote Control stops with it. The files and the conversation stay on the disk, so you can start the machine again and resume the conversation. Before a long wait, turn on Keep awake for that machine.

When you lose signal

Remote Control copes because the session never lived on the phone: Anthropic's docs say the transcript is stored on its servers while connected, which keeps your devices in sync. The weak spot is the computer's own connection. In server mode (claude remote-control), if the computer can't reach the network for about ten minutes, the process exits and you start it again. That matters most on home broadband.

Knowing when the agent needs you

• Remote Control pushes to your phone. Run /config in Claude Code and turn on "Push when actions required" for permission prompts and questions, or "Push when Claude decides" for news such as a finished task. It works on all three setups, wherever Claude Code is signed in to a claude.ai plan. It skips the push while you're typing in the connected terminal.
• A hook works without the app. Claude Code's Notification hook fires once a permission prompt has waited about six seconds, and about a minute after Claude finishes if you haven't typed since. Point it at any push service. Granda's hook catches Claude's questions and sends them through Poke to his phone.
• Plain doom coding has neither. You open Termius and look.

What each costs in a month

Prices as of 2026-11-02, before tax:

The Mac mini arithmetic: Apple lists the current model, with the M6 chip, at 4 W idle and 70 W at most, and the U.S. average residential price was 18.16 cents per kWh in August 2026. Idle for a month (730 hours) is 2.92 kWh, about $0.53; flat out, 51.1 kWh, about $9.28. The cheapest Mac mini on Apple's US store is $899, or $24.97 a month over three years.

Hetzner's price includes an IPv4 address but not VAT, and it bills a server until you delete it. shedix's M is $0.16 an hour, and an S (2 vCPU, 4 GB) is $0.06, so the same months on an S come to $5.28 and $10.56. shedix is in a private beta and charges nothing yet; these are its published rates.

The tools in between cost little: Tailscale's Personal plan and Termius's Starter plan are free. The bill all three share is the agent. Remote Control needs a Claude Pro plan or above, not an API key, and Pro is $20 a month, or $17 a month paid yearly.

Which to choose

• You have a computer at home that can stay on: setup 1 costs next to nothing. Add tmux, and check its power settings.
• You steer agents from your phone most of the day: an always-on VPS gives you a fixed bill and nothing at home to look after.
• You check in now and then: an hourly VM costs least for scattered use, as long as you turn on Keep awake for long waits.

If the third fits, the pricing page lists every size, and the Remote Control post shows how a shedix machine appears in the Claude app.