Approve AI agent permissions without switching windows
Coding agents are careful by design: before an edit or a risky command, they stop and ask. That's the right default — but it turns you into a full-time approval clerk, alt-tabbing into terminals a dozen times an hour. Here's how to keep the safety without the window-switching.
Why agents keep stopping to ask
Permission prompts are the safety layer of agentic coding: file edits, shell commands, and network access wait for a human yes. You can loosen the policy — many developers auto-allow reads and keep prompts for writes — but some prompts always remain, because they should. The problem isn't that agents ask. It's where they ask: buried in a terminal that's behind your editor, on another Space, or on the other monitor. The agent stops instantly; you find out minutes later.
What the alt-tab actually costs
Each approval is two context switches — out of your work, then back — plus the refocus tax after each one. A dozen approvals an hour means your deep-work blocks never exceed a few minutes. And the mirror cost on the agent side is worse: an unnoticed prompt means the agent idles at zero progress while its rate-limit window keeps rolling (see how the windows work). Two resources burn at once, and neither shows up anywhere until you go looking.
Approving from the notch
Vibe Island renders the request in your Mac's notch the moment an agent asks: which tool wants to run, on what, from which session — with Allow and Deny one click away. A distinct sound tells you a request arrived, so you don't need to watch for it. Multiple-choice questions appear as clickable options, and when Claude Code proposes a plan, the whole plan renders with Markdown for you to approve or push back on. You never leave the app you're in; the agent never waits longer than it takes you to glance up.
Which agents support it
Two-way interaction — approvals and question answering from the panel — works with Claude Code, OpenCode, and MiMoCode today. The rest of the supported agents (Codex, Gemini CLI, Cursor, and more) are one-way: live status and notifications in the notch, with a one-click jump to the exact terminal — window, tab, or split pane, across 20+ terminals — when a prompt needs you there. In practice that still removes the hunting, which is most of the pain.
Common questions
Is approving from the notch as safe as approving in the terminal?
It's the same decision, delivered to the same agent — the notch just renders the request where you can see it. You still see which tool wants to run and on what. Everything stays local on your Mac; no cloud service sits in the approval path.
Can I answer the agent's questions too, or only approve tools?
Both. When Claude Code asks a multiple-choice question, the options appear in the panel and you answer with a click. Plan reviews work the same way — the full plan renders with Markdown and you approve it or send feedback.
What happens if several agents ask for permission at the same time?
Requests queue in the panel, each labeled with its agent and session, so you handle them one by one instead of hunting through terminals to find who's blocked.
What about agents that don't support notch approval?
They still show live status and completion notifications, and clicking their session jumps to the exact terminal tab or pane so you can respond there. Two-way approval works with Claude Code, OpenCode, and MiMoCode today; support expands as agents expose the hooks for it.
Keep going
Approvals are half the story — the other half is knowing what your fleet is doing between them: how to monitor multiple agents at once. Or go straight to the product: multi-agent monitoring and the session tracker overview.
Vibe Island · macOS 14+
Your agents ask. You glance up. Work continues.
Download Free Trial2-day free trial · all features included · no card required