Docs

Permissions

Before an agent edits a file, runs a command, writes to another session or starts one, it asks. You answer from any device, or let a harness do some of it without asking.

Requests

A request shows on every paired device with the agent's own options, usually Allow, Always and Reject. The first answer wins, from whichever device; the others see the card close. See the chat for how it looks.

Some agents never ask before they write a file or run a command. For those (Grok, among the built-in ones), the daemon asks on their behalf before each write and each command.

Three modes per harness

Each harness has a mode on each computer, which covers all its sessions there. Set it from the computer, or with /permissions in any session of that harness:

$ slashcode agent permissions claude editsclaude: edits (file edits are accepted, everything else goes to your devices)
  • ask (Ask in the app): every request comes to your devices. The default.
  • edits (Auto-edit): file edits inside the session's folder go through by themselves; everything else asks.
  • all (Auto): every request goes through by itself.

The setting lives on the computer, in ~/.slashcode/permissions.json, and applies from the next request without a restart. slashcode agents shows accepts edits or accepts all beside a harness that is not on ask.

What edits lets through

Under edits, the daemon accepts a request on its own only when all of this holds:

  • The agent says the tool is an edit, a delete or a move.
  • It names at least one file.
  • Every file it names lies inside the session's folder, once .. and symbolic links are resolved.
  • None of them lies in the folder's .git.

Everything else still asks: commands, web fetches, files outside the folder, anything in .git, and requests that name no file. A write to a shell profile, a launch agent or a git hook runs code later, so it is not an edit here: that is what all is for.

Never “always”

When the mode answers for you, it picks the agent's Allow option, the one that holds for this request only, and never an “always” option. The agent keeps no rule from an automatic answer, so changing the mode back to ask takes effect at once. If the agent offers no such option, the request comes to your devices.

Always stays yours to choose when you answer yourself.

Written down all the same

A request the mode accepted still appears in the session, with the option chosen, answered by the agent's permission mode. Scroll back and you see everything that went through, and who let it.

Claude's plan mode

Claude Code has two modes of its own, picked from the model button or /mode: Default and Plan, which plans first and changes nothing until you approve.

When the plan is ready, Claude asks Ready to code? with Yes, go ahead and No, keep planning. Once you say yes, it works under the harness's permission mode as usual. That question is a request like the others: under ask and edits it comes to you, under all it is answered yes on its own.

Sessions, notes and automations

An agent listing your other sessions, waiting for one or reading automations never asks. Its other Slashcode tools ask under the harness's mode: Send a message to other sessions, Start a session, Notify you and Change automations. None of them is a file edit, so only all lets them through without you.

Shell commands

A message starting with ! is run by the daemon itself, with your login shell in the session's folder, without going through the agent or any permission request. It is a full shell from your phone, so it is off until someone at the computer turns it on:

$ slashcode shell onShell commands are on: paired devices can run `!command` on this machine.$ slashcode shell off

The switch is read at each command. While it is off, a ! message ends with a note saying how to turn it on.

A message from another session that starts with ! is never run: it reaches the agent as words, like any other message.