Agent Permissions (Autonomous or Manual)
- Desktop app
- All plans
What it does
Every agent CLI that offers a bypass has a permission mode in AgentsRoom: Autonomous launches the agent with that CLI's own bypass flag, Manual leaves the CLI asking before its actions. An agent running without you (backlog ticket, scheduled task, team step) is only useful if it does not stop on a question nobody is there to answer. Every action stays visible in the terminal, and you can stop the agent at any time. Full page: https://agentsroom.dev/features/agent-permissions
Video (1 min 11): Agent permissions, shown in the real app. Claude Code started by hand with --dangerously-skip-permissions in a new folder stops on "Yes, I trust this folder"; in Settings, AI providers, Claude, Agent permissions is on Autonomous with its four switches; an agent AgentsRoom launched in another new folder writes a README and lists the files without a single question.
Where to find it
- First launch of a CLI: the Agent permissions question, Autonomous (marked Recommended) or Manual. Never asked = Autonomous.
- For a whole CLI: Settings > AI providers, select the CLI, then Agent permissions. Applies to new sessions of that CLI. The Settings button in the footer opens a small menu: its AI providers entry opens Settings on that page, Claude selected. The Agent permissions entry of the left menu scrolls to the Autonomous / Manual choice; the four Claude switches sit right below it: Accept the Bypass warning for me, Skip Claude's account email check, Answer Claude's deletion check, Trust the folders I launch agents in.
- For one agent: Edit agent > Advanced options > Agent permissions: Inherited, Autonomous or Manual.
- For a scheduled task: Permission mode in the task's Advanced section, same three values.
How to use it
- Choose once per CLI at its first launch, or later in Settings > AI providers.
- Keep a review agent on Manual while the others run freely, with the per-agent override.
- Your own flags go in CLI options (provider tab or one agent); in Autonomous, AgentsRoom still adds the bypass flag you did not set yourself. To run Claude Code in its auto mode rather than bypass, choose Manual and add
--permission-mode auto.
Settings
agentPermissionMode(global, Settings > AI providers): provider id toautonomousormanual.preAcceptBypassWarning(global, Claude tab, on): accepts Claude Code's Bypass Permissions warning before an autonomous launch.preAcceptFolderTrust(global, Claude tab, on): answers "trust this folder?" for the folder the agent launches in. The home folder always asks.skipClaudeRemovalCheck(global, Claude tab, on): answers Claude's deletion check for autonomous agents, except on a disk root, a system, home or project folder.skipClaudeEmailCheck(global, Claude tab, on): turns off Claude's account email check for autonomous agents, for their session only.
Agent tools (MCP)
settings_get/settings_set: read or change the keys above (scope: "global").
Providers
- Flag passed in Autonomous: Claude
--dangerously-skip-permissions, Codex--sandbox danger-full-access -a never, Antigravity--dangerously-skip-permissions, Grok--always-approve, Mistral Vibe--agent auto-approve --trust, Kimi Code--auto, OpenCode--auto, Amp--dangerously-allow-all, oh-my-pi--yolo, GitHub Copilot--allow-all-tools, Devin--permission-mode dangerous, Cursor--force, Command Code and Muse Code--yolo, ZCode--mode yolo. - OpenCode's
--autoapproves what you did not explicitly deny; Manual keeps its own prompts. - oh-my-pi runs in yolo by its own default, so Manual passes
--approval-mode always-ask. - GitHub Copilot has no Manual flag: AgentsRoom applies its flag unless you set your own CLI options.
- MiniMax Code has no flag: AgentsRoom writes the mode into its config file, which also applies to
mcodeoutside AgentsRoom. - Aider always auto-confirms, Pi never asks for a tool approval by design, Freebuff has no permission flag.
Mobile
The first-launch Agent permissions question also opens as a sheet on the phone, so an agent launched from the phone does not wait for an answer on the computer. Changing the mode per CLI is done on the desktop.
Limits
- Autonomous is not a sandbox: an agent can delete, push or send what its command allows.
- What AgentsRoom leaves alone: your own
permissions.askrules and hooks answering ask, a managed policy that disables bypass mode, deny rules, and real questions from the agent itself. - When an agent does stop, it gets the red halo, the inbox (Cmd/Ctrl+Shift+I) and a push notification on the phone, where you can answer.
Common questions
- I skipped permissions and Claude Code still stopped. Why? Bypass mode does not skip the launch warning, the folder trust question, some deletions or the account email check. The four Claude tab switches above answer them, all on by default.
- Does OpenCode ask for every action? Not in Autonomous: AgentsRoom passes
--auto. In Manual it keeps its own prompts. - How do I find the setting? It picks Autonomous or Manual for each CLI: click Settings in the footer, then AI providers, select the CLI, and look at Agent permissions. Video: https://youtu.be/zEk-QkzPwFk
- Is there a video that shows it? Yes, about a minute, filmed in the real app: https://youtu.be/zEk-QkzPwFk
- Can one agent ask while the others do not? Yes, set that agent to Manual in Edit agent.
Related
- Multi-Provider Support: every CLI AgentsRoom runs and how each one differs.
- Needs Input Inbox: where a stopped agent shows up.