Is it safe to give Claude control of your Mac?
The short answer: it is as safe as the limits around it. An AI assistant that can run commands and click on your Mac can also make mistakes there, and it can be misled by text it reads. Remotehand does not make that risk disappear. It gives you the controls to decide how much you allow, to see what happened, and to stop it at once.
What can go wrong
- Mistakes. The assistant misunderstands and runs the wrong command, in the wrong folder.
- Misleading content. A web page, e-mail or file contains text written to look like an instruction (“prompt injection”), and the assistant acts on it.
- Too much access for too long. You allow everything once and forget to turn it off.
What Remotehand does about it
- Modes. In Off every request is refused. In Watch only the assistant can take screenshots and read, and nothing else. Remote work allows commands, files, clicks and typing, runs for the hours you set and ends by itself.
- Approvals. Risky actions, such as
sudo, deleting files, installing software, changing system settings, payments or sending mail, wait until you approve that one exact action. A request you do not answer expires after 10 minutes. - A kill switch. ⌃⌥⌘K sets the mode to Off, stops running tasks and jobs and denies pending approvals.
- Things it never does. Remotehand does not let the assistant change, stop or delete Remotehand itself, or export your keys. Not in Remote work and not inside a trust window.
- A record. Every action, including refused ones, is written to an activity log on your Mac. Each entry is chained to the one before, so a changed or removed entry shows.
- Screen content is treated as data. What the assistant sees in a screenshot or reads in a file is passed on as content, not as an instruction from you.
What it does not do
- It does not make the assistant right. An approved command runs exactly as written, so read it before you approve.
- The trust window removes the approval step on purpose. Inside it, risky actions run without asking. Use it for work you would be comfortable leaving unattended, and for as short a time as you need.
- Rules that sort commands into safe and risky cannot foresee everything. A command that looks harmless can still do something you did not want.
- It is not a sandbox. In Remote work the assistant acts with your user account’s rights.
Sensible habits
- Start in Watch only. Switch to Remote work when there is a job to do, for the hours it needs.
- Keep backups, as you would anyway. Time Machine is enough.
- Leave the trust window off unless a task really needs it.
- Look at the activity log after unattended work.
- Do not point the assistant at content you do not trust while risky actions are allowed without approval.
Privacy
Remotehand has no account and no telemetry. Your activity log, screenshots and keys stay on your Mac. Details are on the Privacy page.
If you want to try it with these limits in place: how to let Claude Desktop run terminal commands on your Mac.