DEV Community

herval
herval

Posted on AI-assisted

All your Claude Code sessions and permission prompts in one Mac app, and what a hook really does

Yes, there are Mac apps that show all your Claude Code sessions in one place and let you approve permission prompts. They split into two kinds. Notch and menu bar apps, such as Vibe Island, Coucou, Openbar and AgentBar, hook into the sessions in your terminals so you can allow or deny without switching tabs. Chat apps, such as littleguys, show every session as a thread you read and reply to.

Disclosure: I make littleguys, the second kind. So the useful part of this post is further down: what actually happens inside Claude Code when one of these apps answers a prompt. If you're building one, it may save you some time.

The apps

App What it is Answers prompts from Cost
Vibe Island Notch app for about 30 agents, over SSH too Terminal sessions $19.99 once, closed source
Coucou Notch app, plus your iPhone's Lock Screen Terminal sessions Free, MIT
Openbar Notch app that also takes typed or spoken replies Terminal sessions Free, GPL-3.0
Hushtop Notch app with usage limits Terminal sessions $9.99 after a 7-day trial
Vibe Notch The smallest one, Claude Code only (formerly Claude Island) Terminal sessions Free, Apache-2.0
AgentBar One keyboard-driven queue, with rules you write Terminal sessions Free, MIT
littleguys WhatsApp-style chat app: every session is a thread Terminal and background sessions Free, MIT

The short version: if you like your terminals and want prompts to stop hiding in them, get a notch app. If you want to read what your sessions are doing and answer them, not only approve them, you want a chat app.

They all use the same hook

Claude Code has a PermissionRequest hook. It fires when a tool call needs a decision, and it can answer:

{
  "hookSpecificOutput": {
    "hookEventName": "PermissionRequest",
    "decision": { "behavior": "allow" }
  }
}
Enter fullscreen mode Exit fullscreen mode

Print that and the prompt is answered. An empty answer leaves it to Claude Code. The docs cover this much. What happens when a person and a hook are both answering is another story, and that's the whole product.

What really happens

We tested this against Claude Code 2.1.318 in tmux, in manual and auto mode. (By "we" I mean Claude Code, poking at a second Claude Code.) None of this is promised by the docs, so test it again when Claude Code updates.

The dialog doesn't wait for the hook. They race. The dialog was on screen 3 to 56 ms after the hook started.

What happens Result
You say Yes in the terminal The tool runs at once. The hook is not killed or signalled. If it's waiting on curl, it sits there until its timeout, and the spinner reads "running PermissionRequest hook". A late reply is ignored.
You say No or press Esc The turn is interrupted. SIGTERM reaches the whole tree, then a hard kill. The connection closes within about 0.2 s.
The hook answers first The dialog closes. Allow runs the tool. Deny gives the model your message and the turn goes on.
The hook hits its timeout SIGTERM at exactly the limit, then a hard kill. The dialog stays up and still works.

The first row is the trap. If your app shows an approval card and the person says Yes in the terminal, nothing tells you. The obvious approach leaves a stale card on screen, and a curl running, for as long as your hook timeout, which can be a day.

You can't get this from the transcript. Nothing is written while an approved tool runs: it stayed the same size through a 12-second command, and the tool_use line isn't reliably on disk when the hook fires. PostToolUse doesn't help either, since it only fires after the tool finishes.

What works is claude agents. While the dialog is open, the listing shows the session as:

{ "status": "waiting", "waitingFor": "permission prompt" }
Enter fullscreen mode Exit fullscreen mode

In a 10 ms sample, that was set 18 ms before the hook's request arrived, and cleared 14 ms after the keypress. It held after /clear, after --resume and in auto mode, with the same session id the hooks report.

One more useful number: Claude Code's own permission_prompt notification fired 6.00 seconds after the dialog opened, and not at all if you answered sooner. So that's when Claude Code itself decides you aren't looking. If you wait those 6 seconds before showing anything, someone at the keyboard never gets bothered.

How littleguys handles it

littleguys answers prompts for background sessions, the ones you start from littleguys or with claude --bg, and for terminal sessions, using everything above. A terminal prompt is held for 6 seconds, shown only if claude agents says the dialog is still open, and cleared when the listing stops saying waiting, the hook dies, or the turn ends. Whichever side answers first counts. If Claude Code ever stops reporting waitingFor, the card doesn't show and the prompt stays in the terminal, which is the safe way to fail.

It still doesn't answer AskUserQuestion, since those need an answer picked, and it only follows Claude Code on the same computer.

Drafted with AI help from littleguys' code and pull requests, and each app's own pages. I checked it and I stand behind it.

Top comments (0)