Claude Code Remote Control on a VPS: run Claude Code 24/7 and drive it from your phone

You start a long task in Claude Code, close the laptop, and everything stops. This is my daily setup with Claude Code Remote Control: one VPS, about ten background sessions, and the Claude app in my pocket. Plan for an hour.

Why run Claude Code on a VPS?

A Claude Code session lives on the machine that starts it. On a laptop, it stops when the laptop shuts down. On a VPS, a computer you rent in a data centre that stays on all the time, it works overnight, and you pick it up on your phone in the morning.

With no server at all, cloud sessions on claude.ai/code↗ (opens in a new tab) run at Anthropic. A VPS keeps your own tools and files for a fixed price.

What is Claude Code Remote Control?

Claude Code Remote Control connects claude.ai/code or the Claude app for iOS and Android to a session running on your machine. The work stays on that machine, with your files and tools. Your phone becomes a window and a keyboard, kept in sync with the terminal.

The session only makes outbound requests to Anthropic and opens no port, the numbered doors through which a server accepts connections. While it is connected, the conversation is stored at Anthropic to stay in sync. With /config, it can also notify you when a task finishes or needs your decision.

How do you install Claude Code on a VPS?

Connect to the VPS over SSH, the secure way to reach a server from your terminal. Run Anthropic's install command, then type claude. The server has no browser, so Claude Code shows a link to open on your computer, and you paste the code you get back into the terminal.

1. 1

   Connect to the server

   If SSH runs on another port, add -p and the number.

   Copy

   ssh you@your-vps-address

   Expected result: the server's prompt, for example you@vps:~$

2. 2

   Install Claude Code

   Copy

   curl -fsSL https://claude.ai/install.sh | bash
   claude --version

   Expected result: a version number, for example 2.1.309 (Claude Code)

3. 3

   Sign in

   Run claude, press c to copy the link, open it on your computer, then paste the code where the terminal says "Paste code here if prompted".

   Copy

   claude

   Expected result: Login successful

   Check: /status shows your account and plan.

Do not use `claude setup-token`: the documentation says that token cannot start Remote Control sessions.

How do you drive Claude Code from your phone?

On the VPS, go to the project folder and run claude remote-control. Answer yes to the two confirmations, then press the space bar to show a QR code. Scan it with your phone and the session opens in the Claude app. It also shows up in the session list on claude.ai/code.

1. 1

   Start the Remote Control server

   Copy

   cd ~/my-project
   claude remote-control

   Expected result: "Trust <folder>? [y/N]", then "Enable Remote Control? (y/n)" the first time. Answer y.

2. 2

   Open the session on your phone

   Press the space bar and scan the QR code.

   Check: type "list the files in this folder" on your phone; the VPS terminal shows the same conversation.

This mode accepts 32 sessions at once by default, all in the same folder. For one git working copy per session, add --spawn worktree.

How do you keep Claude Code running 24/7?

A Remote Control session lives as long as its claude process runs: close the SSH connection that started it and it stops. So start it inside tmux, which keeps a terminal open after you leave. For other tasks, claude --bg starts a background session that survives logging out.

1. 1

   Open a tmux session and start Remote Control

   Copy

   tmux new -s claude
   claude remote-control

   Expected result: the session link and QR code.

2. 2

   Detach

   Press Ctrl+B, then D. You can close SSH.

   Check: the session still answers from your phone.

3. 3

   Come back later

   Copy

   tmux attach -t claude

For tasks I start on the server myself, I use claude --bg. A supervisor runs them with no terminal open: as I write this, it has been up for 24 hours with 10 sessions.

1. 1

   Start a task in the background

   Copy

   claude --bg "fix the failing test in the billing folder"

   Expected result: a short id.

2. 2

   Follow your sessions

   Copy

   claude agents

   Expected result: sessions running, waiting for your answer, or finished.

Three limits:

• A server reboot stops tmux and background sessions. Start them again.
• After about 10 minutes without network, claude remote-control exits.
• A session stops when your login expires. When you see "Your login expires in 3 days", run /login.

How do you lock your VPS down?

Remote Control needs no open port, so close the rest. Sign in with an SSH key, never a password. Keep SSH on a private network such as Tailscale. And check Docker: a container published on every address goes straight past the ufw firewall.

I learned this the hard way. In September 2026, on this VPS, RabbitMQ (a message queue) and KeyDB (an in-memory database, with no password) ran in Docker. ufw blocked everything by default, yet both answered from the internet. Someone created two administrator accounts in RabbitMQ, "miner" and "exploit". There was no mining program on the server, but the door had been open for two weeks.

The cause: Docker adds its own network rules, and they run before ufw's. The fix: publish each port on 127.0.0.1, the address only the machine itself can reach.

services:
  keydb:
    ports: !override
      - "127.0.0.1:6379:6379"
  rabbitmq:
    ports: !override
      - "127.0.0.1:5672:5672"
      - "127.0.0.1:15672:15672"

Without a terminal, start with your host's firewall in its control panel: block all incoming traffic except SSH. It filters before your server, so Docker cannot get around it.

1. 1

   Turn off password login

   In /etc/ssh/sshd_config, set PasswordAuthentication no and PermitRootLogin no. Keep your session open and test from a second terminal.

   Copy

   sudo systemctl restart ssh

   Check: ssh -o PubkeyAuthentication=no you@your-vps-address is refused.

2. 2

   Install Tailscale

   Tailscale builds a private network between your devices. Install it on the VPS and on your computer.

   Copy

   curl -fsSL https://tailscale.com/install.sh | sh
   sudo tailscale up

   Expected result: the VPS shows up in your Tailscale devices.

3. 3

   Keep SSH on Tailscale

   Allow SSH on tailscale0 (replace 22 with your port), then delete the rule that opened it to the internet.

   Copy

   sudo ufw allow in on tailscale0 to any port 22 proto tcp

   Check: ssh works through Tailscale and fails from any other network.

4. 4

   Republish containers on 127.0.0.1

   Add the file above, then recreate the containers.

   Copy

   docker compose up -d --force-recreate
   ss -tln

   Check: your ports show behind 127.0.0.1, not 0.0.0.0.

5. 5

   Test from outside

   From a machine off Tailscale:

   Copy

   nc -zv your-vps-address 6379

   Expected result: the connection fails.

On my VPS, SSH listens on port 2222. Moving the port protects nothing: it only cuts the noise from bots trying port 22.

How do you limit what Claude can break on the server?

Run Claude Code under a user with no administrator rights, or with rights limited to a few commands. Never paste a password or an API key into the conversation. And keep merging code, and anything that touches production, for a human to approve.

My VPS has two users, one without sudo. To change the firewall, the agent got sudo for iptables only, for that job. My agents hold no GitHub token: a separate program keeps the key and opens pull requests, and I merge myself.

Your Claude login is stored in ~/.claude/.credentials.json. Whoever gets into your Linux account uses your subscription, one more reason to lock down SSH.

Why does your weekly limit run out faster?

Each background session draws on your quota on its own, and a long conversation costs more with every message, because Claude rereads all of it. On my VPS, 81% of usage came from conversations over 200,000 tokens, the unit of text Claude reads, about three quarters of a word.

One session left open for twelve days, coordinating 94 subagents, took 61% of a week's quota. Since then:

• one fresh session per task;
• at most two sessions building or testing at once on 8 cores and 23 GB. With three, tests failed for lack of memory.

What about OpenClaw or Hermes on the same VPS?

The same rules apply to any AI agent that runs nonstop, including OpenClaw and Nous Research's Hermes Agent. The agent holds your keys and runs commands, so give it a separate user, publish its Docker services on 127.0.0.1, and keep access to the server on your private network.

OpenClaw on a VPS gets its own article.

Common errors

Frequently asked questions

Keep reading

• Claude skills: the skills I use every day, to give your sessions your own instructions.