> ## Documentation Index > Fetch the complete documentation index at: https://docs.rundock.ai/llms.txt > Use this file to discover all available pages before exploring further. # Permissions > What agents can do without asking, what they ask about, and how to change or take back an answer. Agents act on your machine, so Rundock decides what they can do on their own and asks you about the rest. Three things set the line: the workspace's mode, whether agents are kept inside the workspace, and the answers you have given before. ## Permission cards When an agent wants to do something that needs your say, a card appears in the conversation. It says what the agent wants to do in plain words, and shows the command or the file. * **Allow** lets this one request through. * **Always allow** lets it through and remembers the answer for this workspace. * **Always allow this folder** appears when the request reaches a folder outside the workspace. It adds that folder to your working folders, which covers commands as well as file edits from then on. * **Deny** refuses it, and the agent is told. In Code mode, a command that can't be undone, such as a force push or deleting files git has never saved, gets a card that says why and offers **Allow once** with no Always allow. Rundock asks about those every time; in Notes mode commands ask anyway. A card raised by a [routine](https://docs.rundock.ai/concepts/routines.md) appears in the approvals dock, over whatever screen you have open, named for the routine and its agent. ![Rundock permission card in a conversation: the Content Lead asks to edit a file in a client folder outside the workspace, with Allow, Always allow this folder and Deny.](https://mintcdn.com/rundock/1StrqdLhYQpv9eWl/images/permission-card.png?fit=max\&auto=format\&n=1StrqdLhYQpv9eWl\&q=85\&s=5da1a4e53d74aefc8392d7a9c03f6784) A card for a change outside the workspace, offering to allow the whole folder. ## Notes and Code Each workspace has a mode, set in **Settings**, **Permissions**, under **What are you working on?** * **Notes**, for notes, documents and other files. Agents ask before running commands, and cannot write code files such as `.js`, `.py` or `.sh`. * **Code**, for websites and software. Agents edit code and run everyday development commands without asking. Commands you could get back with tools you already have, such as git, a rebuild or a reinstall, run straight away. Commands that lose something for good ask every time. A few that other people see first, such as pushing to the default branch, ask once and can be remembered. The mode applies to every agent in the workspace: Codex agents' commands follow the same rules as Claude Code's. The block on code files in Notes mode applies to Claude Code agents; a Codex agent's file changes inside the workspace are accepted in either mode. ## Keep agents inside this workspace A separate switch, below the mode, decides whether agents can change files outside the workspace folder. On macOS it uses the system's own sandbox and starts on for a new workspace. Under it, **Folders agents can also change** lists your working folders: name a parent such as `~/Projects` and everything inside it is covered, including folders you create later. See [Keep agents inside this workspace](https://docs.rundock.ai/reference/workspace-structure.md#keep-agents-inside-this-workspace) for every state the switch can show. ## Take an answer back **Tools allowed without asking**, at the foot of **Settings**, **Permissions**, lists every Always allow you have given, in words, including commands you chose to remember in Code mode. Remove one and the card asks again next time. Remove a working folder from the list above it to stop covering that folder; conversations already running keep the folders they started with. ![Rundock Settings open at Permissions: Code selected under What are you working on, the Keep agents inside this workspace switch on with \~/Projects listed as a folder agents can also change, and two entries under Tools allowed without asking, each with a remove control.](https://mintcdn.com/rundock/1StrqdLhYQpv9eWl/images/settings-permissions.png?fit=max\&auto=format\&n=1StrqdLhYQpv9eWl\&q=85\&s=0ee5d190941e1bb61a5fd7d085bab79a) Settings, Permissions: Code mode, agents kept inside the workspace with one working folder, and two remembered answers that can be removed. ## What always asks Some things ask on every occasion, whatever you have allowed: any access to your Claude account credentials, and any change to your global instructions (`~/.claude/CLAUDE.md` with its `rules` and `output-styles` folders, and the Codex equivalents, `~/.codex/AGENTS.md` and its `rules` folder) or to a workspace's own permission files. Reading the instructions doesn't ask. A change an agent makes to the permission files is put back first, then you are asked whether to keep it. ## Where to next > **[Workspace structure](https://docs.rundock.ai/reference/workspace-structure.md#workspace-mode)** > > Where the mode, the switch and working folders are stored. > **[Runtimes](https://docs.rundock.ai/concepts/runtimes.md#permissions-and-sandboxing)** > > How Claude Code and Codex differ in their sandboxes. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.