How safe is approving Claude Code from a phone?
Updated
Quick answer
Approving Claude Code from your phone is safe when three things hold: only a phone you paired can answer, the phone can only answer a prompt Claude Code was about to show, and your deny rules win whatever is tapped. What remains is human: a command approved without reading it, an unlocked phone in someone else’s hands, and project details shown on a lock screen.
What a remote approval can and cannot do
Remote approval runs through a PermissionRequest hook. Claude Code runs it when it is about to ask you for a permission, passes it the tool and its input, such as the shell command, and accepts allow or deny as the answer. From there, the limits are Claude Code’s, not the phone’s:
- It only answers a question Claude Code asked. The hook fires when a permission prompt is about to appear, not before every action. No prompt, nothing to approve.
- Deny rules win. A hook that returns allow doesn’t get past the deny rules in your settings, and deny rules hold in every permission mode,
bypassPermissionsincluded. Rules are checked in the order deny, then ask, then allow. - Some prompts can’t be skipped by a hook’s allow, such as reads from network paths and MCP tools marked as requiring user interaction.
- The mode decides how much is asked. In
bypassPermissionsnothing is asked, and inautomost actions run after background safety checks. Remote approval matters in the modes that ask, such as manual mode andacceptEdits.
One thing a phone can do beyond allow and deny is send text. With Monish, a reply to a finished turn becomes Claude’s next instruction, exactly as if you had typed it. Whoever can answer from your phone can therefore also tell Claude what to do next. That is the reason for the next section.
Who can answer
The first control is who holds the other end. With Monish, a computer reaches only the iPhone it was paired with, through a 6-digit code shown on that iPhone and run on the computer. The code is good for 10 minutes, and the app’s key is made at random on the iPhone, with no account and no password to steal.
The second control is the phone itself. Monish doesn’t ask for Face ID before Allow: the buttons on the lock screen, on a notification or on a session card, work as they are, so anyone holding your iPhone can tap them. Keep it with you, and see below for what shows while it is locked. If a phone is lost, run monish uninstall on each computer: it removes the hooks and unpairs it, and nothing reaches the old phone after that.
Other setups have their own version of this question. A chat bot that relays approvals is as safe as the list of people allowed to reply to it: Anthropic’s channel docs warn that anyone who can reply through a channel can approve or deny tool use in your session. An ntfy topic is protected by its name alone.
Read before you tap
The risk that no system removes is approving without reading. A lock screen invites quick taps, so read the request the way you would in the terminal:
- The whole command, not its first words. Chains with
&&, pipes intosh, and anything after a;are where surprises hide. Monish shows a long command in full behind Show full command. - The project. Every request names it. A request from a project you weren’t expecting is a reason to deny.
- What it touches: deletions, files outside the project, network calls, credentials, anything that deploys or publishes.
When in doubt, deny, and say why if you like: the reason goes back to Claude, which can propose something else. Or don’t answer at all. Nothing is decided for you: once the answer window is over, 10 minutes by default, Claude Code asks you in the terminal.
Then make the phone’s job smaller. Allow rules in your settings pre-approve the routine commands you would approve anyway, such as your test runner, and an ask rule keeps a prompt for the dangerous ones even when a broader allow rule matches. Fewer, rarer requests are easier to read carefully. From the phone, Always allow this command does the same for one exact command in one project; the list lives on the computer’s page in the app, where each line can be removed.
What leaves your computer
A remote approval means part of the request travels. With Monish, it is what the notification needs to be decided: the project name, Claude’s message or question, the command or file the permission is about, and, for the session card, the step Claude is on, such as Editing session.ts. For a file edit, up to 12 of the changed lines too, with values that look like secrets hidden and nothing from files like .env or private keys. Never a whole file. It travels over HTTPS. It is not end-to-end encrypted, so our server can read it while it holds it.
Claude’s messages can quote error output or paths. If a project handles secrets, keep them out of commands Claude runs in the open, such as by reading them from the environment rather than pasting them into a prompt.
What shows on the lock screen
A lock screen is readable by anyone who glances at it. iOS lets you choose, under Settings › Notifications › Show Previews, or for Monish alone under Settings › Notifications › Monish, whether notification text shows always, only when unlocked, or never. With When Unlocked, a request’s notification hides its command or question until Face ID or your passcode unlocks the iPhone. That setting covers notifications, not Live Activities: while a request waits, its card on the lock screen shows the project and the command or question, with Allow and Deny for a permission, until you answer or the request closes. Monish has no setting to hide their details. If that is too much, turn off Live Activities in the app’s Settings or in the iPhone’s Settings › Monish: requests then arrive as notifications only, which Show Previews covers. While a session runs, its card also shows the project, the step Claude is on and how long it has been working; the same switch removes it.
What Monish keeps
- No identity: no name, no email, no phone number.
- Notifications: kept 48 hours on our server so the app can catch up on one it missed, then deleted.
- Answers: kept 24 hours.
- Your feed lives on your iPhone.
To remove everything, run monish uninstall on each computer, then Delete account in the app, which erases everything on the iPhone and on our server. The privacy policy has the details.
Questions
- Can a remote approval get past my deny rules?
- No. A hook that returns allow doesn’t bypass the deny rules in your settings, and deny rules hold in every permission mode, bypassPermissions included.
- What if I lose my iPhone?
- Run
monish uninstallon each paired computer: it removes the hooks and unpairs the computer, so nothing reaches the old phone. Requests already on it expire once the answer window is over. - Is the request encrypted?
- It travels over HTTPS. It is not end-to-end encrypted: our server can read it while it keeps it, 48 hours for a notification and 24 hours for an answer.
- Do I need remote approval in auto or bypass mode?
- Hardly. In
bypassPermissionsmode Claude Code asks for nothing, and inautomode most actions run after background checks. Your deny rules still block in every mode.
More guides
- Approve Claude Code permissions from your iPhoneHow a hook hands your Allow or Deny back to Claude Code.
- The Claude Code PermissionRequest hookAnswer a permission prompt from a script: input, output, limits.
- Claude Code permission modes, and fewer promptsManual, auto, plan and the rest, and rules that cut the prompts you don’t need.
- Long Claude Code tasks: how to step awayWhy long tasks stall, and how to keep one moving while you are out.
New to Monish? The home page shows what lands on your iPhone, and all guides are listed here.
Get Monish on your iPhone
Free, no account. Your 6-digit code is waiting in the app.