claude-controls-mac

Here’s a full step-by-step guide on how to turn your spare Mac into an always-on machine Claude Code can fully control, with computer use enabled. You’ll be able to talk to it from your phone through the Claude app, or from your main Mac over SSH.

In case you’re reading this on GitHub Pages, here’s the repo version.

Why do this?

I wanted to create a separate environment Claude Code can control on its own, so I can delegate tasks I don’t necessarily want to run on my own machine - certain types of research tasks, and development tasks.

Claude Code, especially with the --dangerously-skip-permissions flag on, carries inherent risk when run on your main machine. You can eliminate / mitigate these risks by creating a separate environment on your spare Mac with everything it needs to have access to.

It has an added bonus of being able to talk to Claude Code anytime, anywhere from your phone. I’ve personally found it really useful because I often prefer to talk to Claude Code instead of regular Claude on the mobile app - Claude Code is often more capable.

The following guide assumes you have your main Mac as well as a spare Mac you can set up for this, but you should be able to take inspiration from it and apply it to any combination of two machines.

Why this setup?

First, let’s quickly address a few questions you might have.

Why not run it in a container?

I’m a big proponent of running it in a container - I even built an entire environment for doing so conveniently. However, I’ve found it has a few limitations. First, it still runs on your main machine, so it’s not completely separated. For example, network requests it sends still go through your main machine.

Second, there are limitations to the container’s capabilities. For example, I wanted my agent to be able to run Unity for game development, and there’s no easy way to do that in a container. The same goes for any other app that’s only available on a Mac - you won’t have access to it. That’s especially relevant if you want Claude Code to control these apps through computer use - clicking, dragging, and so on.

Why not use something like OpenClaw?

I personally like having access to the full, latest features of Claude Code. I also like being able to control it from the Claude app - I’ve found it really convenient. And you get to use your Claude subscription usage if you happen to have one, which is an added bonus.

At the end of the day, running an agent with broad permissions is safer on a machine that has nothing to lose - but you get the benefit of being able to use a full Mac instead of a container. The approach here:

• Use an old/spare Mac, not your main one.
• Create a fresh local account with no personal data and no Apple ID signed in, so the agent has nothing sensitive to reach.
• Drive it over SSH from your main Mac on your local network, and control it from your phone.

What you’ll need

• A spare Mac (the target).
• Your everyday Mac (the source), on the same Wi-Fi.

1. Start fresh on the target Mac

Wipe it first (if it has any personal data)

You’ll be giving the agent full access to this machine, so it can reach anything stored on it. If there’s existing data you don’t want it to have access to, erase the machine first:

• Macs that support it: System Settings -> General -> Transfer or Reset -> Erase All Content and Settings.
• Older Intel Macs: restart into Recovery (hold Cmd-R at boot), use Disk Utility to erase the internal drive, then reinstall macOS.

Optionally update to the latest macOS afterward (System Settings -> General -> Software Update).

Create a fresh, isolated account

• Create a new local user account (System Settings -> Users & Groups).
• I recommend not signing into an Apple ID. Skip it during setup.

Make the account an admin (if you haven’t already)

The account needs admin rights or sudo will refuse to run.

• System Settings -> Users & Groups -> set the account to Allow this user to administer this computer.
• If you ever need to repair it from another admin account: sudo dseditgroup -o edit -a <user> -t user admin

2. Enable Remote Login (SSH) on the target Mac

On the target, turn on SSH so the source Mac can connect:

sudo systemsetup -setremotelogin on

If the command fails with Turning Remote Login on or off requires Full Disk Access privileges, give your terminal app Full Disk Access first:

• System Settings -> Privacy & Security -> Full Disk Access.
• Click +, then in the file picker go to Applications -> Utilities -> Terminal and add it.
• Quit and reopen the terminal, then rerun the command.

3. Passwordless sudo for the target account

This is so the agent (and your SSH commands) can run admin tasks without a password prompt each time. Run this once on the target. It asks for the login password this one time:

echo "<user> ALL=(ALL) NOPASSWD: ALL" | sudo tee /etc/sudoers.d/<user>-nopasswd >/dev/null
sudo chmod 440 /etc/sudoers.d/<user>-nopasswd
sudo visudo -cf /etc/sudoers.d/<user>-nopasswd   # validate - must print 'parsed OK'

This creates a small rule file telling the Mac that <user> can run sudo without a password prompt:

• line 1 writes the rule into /etc/sudoers.d/.
• line 2 makes it read-only - sudo ignores the file otherwise.
• line 3 validates the syntax; a typo in a sudoers file can lock you out of sudo entirely, so it must print parsed OK.

After this, sudo runs with no prompt - test with sudo -n true, which succeeds silently if passwordless sudo works.

4. Find the target’s address (hostname or IP)

You can reach the target by either a hostname or an IP. I recommend using the hostname: it stays the same, while the IP can change.

Hostname (recommended). Run on the target:

scutil --get LocalHostName      # prints the hostname, e.g. MacBook-Pro

Add .local to form the address: <target-host>.local. You can also read it from System Settings -> General -> Sharing, shown as Local hostname.

Give the target a unique name. Each Mac needs a .local name that’s unique on your network. If two machines share a name, the address can point to the wrong Mac. Make sure the target’s name is unique - rename it if needed:

sudo scutil --set LocalHostName newmacbook   # -> newmacbook.local

IP address (not recommended). Run on the target:

ipconfig getifaddr en0          # e.g. 192.168.1.80

Note that the IP can change after a reboot or after a certain amount of time.

Throughout the rest of this guide, replace <user> with the target account name and <target-host> with the hostname from above, so the address is <user>@<target-host>.local. You could also instead use an IP in place of <target-host>.local.

5. Set up passwordless SSH from the source Mac

On the source Mac, create an SSH key (skip if you already have one):

ssh-keygen -t ed25519

Install your public key on the target. This asks for the target account’s login password once:

ssh-copy-id <user>@<target-host>.local

Test it - this should print the target username with no password prompt:

ssh <user>@<target-host>.local whoami

6. Keep the target awake

By default macOS sleeps after ~10 minutes idle, even when plugged in, which takes it off the network. To make it never sleep, run this on the target (or over SSH from the source):

sudo pmset -c sleep 0          # never system-sleep while plugged in (-c = on charger)
sudo pmset -c disablesleep 1   # also prevents sleep with the lid closed (clamshell)
sudo pmset -c displaysleep 0   # keep the display on too

Verify:

pmset -g | grep -iE 'sleep'

sleep 0, SleepDisabled 1, and displaysleep 0 in the output confirm it worked.

If the machine runs on battery sometimes, use -a instead of -c to apply to all power sources (at the cost of battery drain).

The screen can still lock when the screen saver kicks in. Stop the screen saver from ever starting so it never locks on its own:

defaults -currentHost write com.apple.screensaver idleTime 0

7. Clipboard sync over SSH

macOS ships pbcopy (write clipboard) and pbpaste (read clipboard). Piped over SSH, they can move the clipboard between machines - encrypted, peer-to-peer, no Apple ID or third-party service.

clip.sh wraps this into one command with two subcommands, and adds image support on top of pbcopy/pbpaste (which are text-only). Install it as a script on your PATH on the source Mac, and point it at the target host with IC_BOX (“ic” standing for “isolated claude”):

curl -fsSL https://raw.githubusercontent.com/ykdojo/claude-controls-mac/main/clip.sh -o ~/.local/bin/clip
chmod +x ~/.local/bin/clip
export IC_BOX="<user>@<target-host>.local"   # add to ~/.zshrc

Usage:

• clip send - this Mac’s clipboard → the target (text or image). For an image you can paste it straight into a Claude Code session on the target with Ctrl-V.
• clip get - the target’s clipboard → this Mac (text or image).

8. Install Claude Code on the target Mac

Send the install command over and run it. From the source Mac you can push it straight to the target’s clipboard, or run it remotely. The following command installs a specific version, but you can also install latest or stable if you’d like:

ssh <user>@<target-host>.local 'curl -fsSL https://claude.ai/install.sh | bash -s -- 2.1.201'

The native installer may warn that ~/.local/bin is not on PATH. Fix it on the target by adding it to ~/.zshenv (not ~/.zshrc) - .zshenv is read by every zsh, including non-interactive ones:

ssh <user>@<target-host>.local 'echo '\''export PATH="$HOME/.local/bin:$PATH"'\'' >> ~/.zshenv'

9. Set up an opinionated, Claude Code-friendly environment (optional)

This optional step applies a set of opinionated defaults via setup-claude-env.sh - shell aliases, the DX plugin, settings.json tweaks, the GitHub CLI, and (opt-in) Playwright MCP and yt-dlp. Every item is toggleable; see the full list in claude-env-components.md.

Interactively on the target - shows a checklist of every item (core pre-checked, opt-ins unchecked) so you can pick any combination. Download the script onto the target and run it:

ssh -t <user>@<target-host>.local \
  'curl -fsSL https://raw.githubusercontent.com/ykdojo/claude-controls-mac/main/setup-claude-env.sh -o setup-claude-env.sh && bash setup-claude-env.sh'

Non-interactively - the script goes into this mode if it detects a non-interactive environment or you supply at least one flag. No prompt; installs core only by default, or pick what you want with flags (--yt-dlp, --playwright, --all, --core):

ssh <user>@<target-host>.local \
  'curl -fsSL https://raw.githubusercontent.com/ykdojo/claude-controls-mac/main/setup-claude-env.sh -o setup-claude-env.sh && bash setup-claude-env.sh --all'

The script is idempotent (OK to re-run).

10. Log in to Claude and GitHub

Both logins are interactive, so SSH in:

ssh <user>@<target-host>.local

Then run claude on the target - it drops into the login for your Anthropic account. Follow the prompts (a browser/device-code flow you can finish from a browser on your main Mac).

GitHub - optional, but highly recommended so the agent can work with repos:

gh auth login

If you haven’t installed the GitHub CLI from the previous step, make sure to do so first.

I personally recommend using a separate GitHub account, not your main one, so it doesn’t mess up your main account.

11. Computer use over SSH (optional)

This lets an interactive claude session on the target see (screenshots) and control (mouse/keyboard) its own desktop, driven over SSH.

This doesn’t work out of the box - SSH and macOS’s permission model get in the way, so the setup below exists to route around that.

Why it needs a workaround: macOS gates screen capture and input behind Screen Recording and Accessibility permissions that are tied to the GUI login session, so an SSH process can’t reach the display. Fix: a LaunchAgent keeps a tmux server alive inside the GUI session on a fixed socket; every claude session created there lands on that server and inherits the GUI session, so it can reach the display. You attach over SSH.

Scriptable setup

Run setup-computer-use.sh on the target:

ssh -t <user>@<target-host>.local \
  'curl -fsSL https://raw.githubusercontent.com/ykdojo/claude-controls-mac/main/setup-computer-use.sh -o setup-computer-use.sh && bash setup-computer-use.sh'

This installs the LaunchAgent (persistent tmux server with anchor session cc) and enables the built-in computer-use tool in ~/.claude.json. Requires tmux (brew install tmux) and a Claude Pro or Max plan. Re-runnable; --uninstall to remove.

Use it from your Mac

Install ic.sh (ic = “isolated claude”) on the source Mac:

curl -fsSL https://raw.githubusercontent.com/ykdojo/claude-controls-mac/main/ic.sh -o ~/.local/bin/ic
chmod +x ~/.local/bin/ic
echo 'export IC_BOX="<user>@<target-host>.local"' >> ~/.zshrc   # or edit the default in the script

Each ic spawns its own claude session on the box (run several at once) and attaches. Flags mirror claude, and it adds several more subcommands:

ic               # new claude session
ic -c            # continue the most recent conversation (forwards to: claude -c)
ic -r            # resume picker (forwards to: claude -r)
ic --chrome      # with Claude in Chrome (forwards to: claude --chrome)
ic sh            # a plain shell on the box, no claude (alias: ic shell)
ic vnc           # open Screen Sharing (VNC) to the box (see step 15)
ic rc            # Remote Control: drive the box from your phone (claude remote-control)
ic history       # stored conversations: count, location, recent w/ previews (alias: hist)
ic ls            # list live sessions (state, age, what's running, conversation)
ic attach <id>   # attach a running session (alias: ic a)
ic kill <id>     # kill a session (alias: ic k)
ic kill-all      # kill all sessions
ic kill-except <id> <id> ...   # kill all sessions except the listed ones
ic -h            # help

All ic sessions run with --dangerously-skip-permissions (and ic rc uses --permission-mode bypassPermissions).

Copying text out: sessions run in tmux, and Terminal.app can’t receive the clipboard escape sequences (OSC52) that claude emits, so mouse-selecting a snippet won’t reliably reach your Mac clipboard. Quick workaround: Cmd-A then Cmd-C copies the whole visible screen.

Bonus: let your Mac’s Claude drive the box’s Claude

Because every ic session lives on the box’s tmux server at a fixed socket, a Claude Code session on your source Mac can prompt one directly over SSH - useful for delegating work to the box and checking on it, agent to agent:

# find the session (ic ls prints the ids)
ic ls

# type a prompt into it, then confirm with a second Enter - with longer prompts,
# an Enter sent in the same burst as the text doesn't register as submit and the
# prompt sits in the input box (short prompts usually submit fine)
ssh <user>@<target-host>.local "tmux -S /tmp/cc-tmux.sock send-keys -t <session> 'Switch to main and pull; PR #4 is merged.' Enter"
sleep 2
ssh <user>@<target-host>.local "tmux -S /tmp/cc-tmux.sock send-keys -t <session> Enter"

# read the reply (re-run / poll until the spinner is gone)
ssh <user>@<target-host>.local "tmux -S /tmp/cc-tmux.sock capture-pane -t <session> -p" | tail -30

One-time grants (can’t be scripted)

Screen Recording and Accessibility can only be granted in the GUI, and a human has to do it at the machine (in person or via Screen Sharing) - macOS blocks synthetic clicks on these prompts. On first capture you’ll also Allow a “bypass the window picker” prompt (recurs ~monthly).

The grants go on tmux, not claude. macOS attributes capture/control to the responsible process in the chain, which here is the tmux server (claude runs as its child). So:

1. Grant tmux (/usr/local/bin/tmux, or /opt/homebrew/bin/tmux on Apple Silicon) under both Screen Recording and Accessibility - Screen Recording covers screenshots, Accessibility covers mouse/keyboard control.
2. Restart the tmux server after granting - a running process caches its permission state at launch, so a grant won’t take effect until the server restarts: tmux -S /tmp/cc-tmux.sock kill-server (the LaunchAgent respawns the anchor within seconds). claude sessions started after the restart pick up the new grant.

To make the entries appear in System Settings in the first place, trigger a computer-use action (ic, then ask Claude to “take a screenshot”) - macOS adds tmux to the list (toggled off) so you can switch it on.

Full Disk Access stops the “access data from other apps” prompts. On newer macOS, tmux triggers a separate “tmux would like to access data from other apps” prompt for each app whose data anything inside it touches. Granting tmux Full Disk Access (same Privacy & Security pane) suppresses these prompts entirely: use the + button and add the tmux binary (Cmd-Shift-G to type the path). Restart the tmux server afterward, as above.

12. Install a VPN, or any other app (optional)

I personally like to run a VPN on the box so its traffic goes out separately from my local IP. I use Proton VPN - it has a free option and I’ve been using them for a long time - but of course there are other options.

You can just ask the box’s Claude to do it if you finished walking through the previous step: ic in and say “install Proton VPN”. It’ll download and install the app. The parts it can’t do alone:

• Credentials. Signing in is required even for the free tier, and that’s yours to enter. Send the password over securely with clip send from step 7: copy it on your Mac, run clip send, then paste into the sign-in field on the box.
• macOS permission prompts. The first connect pops a system prompt to allow a VPN / network configuration (and may ask for the Mac password) - approve it at the machine.
• Computer use for the GUI. The app is GUI-only, so driving it relies on step 11 being set up. Once signed in, the agent can connect and switch servers itself.

This general flow works not just for a VPN app - it should work for pretty much any other application too.

13. Control it from your phone

Remote Control lets you drive Claude Code from your phone through the Claude app. There are two ways to use it:

• Run /remote-control (or /rc for short) inside an existing session and follow the instructions - you can then drive that same session from your phone, going back and forth between the two.
• Start a server with claude remote-control, which also lets you start brand new sessions from your phone, not just attach to one you already have open.

If you went through step 11, ic rc starts that server for you in a way that the sessions you spawn from your phone have access to computer use too.

14. Set up Claude in Chrome (optional)

Computer use from step 11 deliberately won’t control a browser - browsers get a restricted tier where Claude can see what’s on screen but can’t click or type in it. The Claude in Chrome extension fills that gap with proper browser control: navigating, clicking, filling forms, reading console logs and network requests. And unlike Playwright MCP, it drives your regular Chrome profile, so the agent can use any logged-in state you set up on the box.

It requires Chrome on the target and a direct Anthropic plan (Pro, Max, Team, or Enterprise).

You can ask the box’s Claude to install Chrome and open the extension’s Chrome Web Store page. The parts it can’t do alone (do these at the machine, in person or via Screen Sharing):

• Click “Add to Chrome”. Optionally, you can pin the extension to the toolbar by going to chrome://extensions.
• Log into your Claude account in the extension. You can send the login info with clip send from step 7.

Then enable it in Claude Code on the target: run /chrome in a session and select “Enabled by default” to have the browser tools in every session, or skip that and use claude --chrome per session (ic --chrome from your source Mac). If the connection doesn’t work, try restarting Chrome once.

To make browser use efficient (element refs from the accessibility tree instead of coordinates, no unrequested screenshots), the environment setup from step 9 adds a “Claude for Chrome” guidance section to ~/.claude/CLAUDE.md on the target. If you skipped that step, re-run the script - it’s idempotent.

If your source Mac’s Chrome also has the extension signed into the same account, Claude Code can connect to either browser. It prompts you to pick when both are connected, but its local-machine detection can be wrong (#74667). To avoid that, you can temporarily uninstall the extension from the source Mac’s Chrome - or just quit Chrome on the source while the box is working.

Note that sessions spawned from your phone via step 13 don’t currently get the browser tools (#74671) - the workaround is to start the session in the terminal and attach with /rc.

15. Enable Screen Sharing on the target Mac (optional)

This lets you see the target’s screen live from the source Mac - and take over its mouse and keyboard - using macOS’s built-in Screen Sharing.

This has to be enabled in the GUI at the machine - since macOS 12.1 it can’t be enabled from the command line.

On the target: System Settings -> General -> Sharing -> Screen Sharing on. If Remote Management is on, the Screen Sharing toggle may be hidden - turn it off first.

Then connect from the source Mac:

open vnc://<user>@<target-host>.local

Log in with the target account’s login password and tick Remember this password in my keychain.

16. Access it from anywhere with Tailscale (optional)

.local names only resolve on your LAN, so everything so far is local-network only (except phone control from step 13, which relays through Anthropic). Tailscale fixes that: it connects your machines with peer-to-peer, end-to-end encrypted WireGuard tunnels, so SSH, ic, clip, and Screen Sharing work from any network with nothing exposed to the public internet. On your home network it takes a direct LAN path, so local use stays as fast as before. And joining the network only gives a device reachability - SSH and Screen Sharing still require your key or password on top.

Install on the target (the Homebrew formula works headless over SSH; sign in at the URL the last command prints - a free account with a social login is fine):

ssh <user>@<target-host>.local 'brew install tailscale'
ssh <user>@<target-host>.local 'sudo brew services start tailscale'
ssh <user>@<target-host>.local 'sudo tailscale up --operator=<user>'

Install on the source, then open the Tailscale app and log in with the same account - devices only see each other within one account’s network:

brew install --cask tailscale-app

With MagicDNS (on by default) the target is reachable by bare hostname from anywhere - the same address, minus .local. Switch IC_BOX in ~/.zshrc, keeping the original name for reference:

export IC_BOX="<user>@<target-host>"          # Tailscale - works remotely too
# export IC_BOX="<user>@<target-host>.local"  # original - local network only

Screen Sharing works the same way: open vnc://<user>@<target-host>.

Recommended, in the admin console: device approval (Settings -> Device management), so a compromised Tailscale login alone can’t add a device to your network, and disable key expiry on the target (Machines -> …), so the box doesn’t drop off the network when its key expires after ~180 days.

To confirm remote access really works, put the source Mac on a different network (e.g. a phone hotspot) and run ic ls.

If you want more Claude Code and agentic coding tips, feel free to subscribe to my newsletter, Agentic Coding with Discipline and Skill.