AgentDeck lives at the top of your screen and keeps watch over Claude Code and Codex. It drops down the moment an agent needs you, so you can approve, answer or reply without hunting for the right terminal.
Hover the top of your screen to peek at what an agent is doing, without switching windows.
The island above is live: hover it, click its buttons, type a reply.
Three agents are running. One has waited ten minutes for a yes. One finished a while ago. One hit its usage limit. You'd only find out by checking every tab.
Or the island tells you, the moment it matters.
01The crew
Each running agent gets its own little character and keeps it for the whole session, so you can tell them apart at a glance. Their faces tell you the rest.
Eyes dart, and it bobs like it's typing. Blue means busy.
psst, they follow your cursor. Click one.
02Features
Stop cycling through terminal tabs to find the one that's been waiting ten minutes for a "yes". AgentDeck brings every question to you, and gets out of the way when nothing needs you.
Working, waiting for you, done, or resting at a usage limit. Click a session to see what it's doing.
How much of your Claude and Codex limits you've used, with a nudge at 80% and 95%.
Multiple-choice questions drop down from the island. Pick one, and the agent carries on.
Which test runner should I set up?
Type its next instruction on the island and it keeps going. Go on, try it.
When an agent's command runs longer than you like, the island tells you when it ends.
npm run e2e
When an agent pushes, AgentDeck can follow its GitHub Actions with your GitHub CLI. Opt-in, off by default.
git push origin feature/rate-limitEvery request and how it ended: who decided, how long it waited. Risky ones allowed in under 2 seconds get flagged.
npm install --save-dev vitestMEDIUMyou · 3.1s
git push --force origin mainCRITICALyou · 6.4s
git reset --hard HEAD~3HIGHyou · 1.2s ⚑
npm testMEDIUMrule · Always
cat .envHIGHterminal · 9m
Step away from the desk. The Android app shows the agent that needs you, and you answer with a tap. Go on, try it.
Settings → Phone → Connect a phone, and scan the code.
Can I run a command?
npm test
03Safety net
Every request gets a risk level (low, medium, high or critical) with the reason in plain words. Risky ones slow you down just enough to think.
Try itType a command an agent might ask to run
A simplified version of AgentDeck's classifier, running in your browser. The real one understands
bash, PowerShell and cmd: quoting, chained commands, $(…), aliases, wrappers and
encoded PowerShell.
Buttons ignore clicks for 400 ms after a card appears, so a click meant for another window can't approve anything.
Closed, crashed or slow: your agents simply ask in their terminal as always. AgentDeck never approves anything on its own when something goes wrong.
The answering key lives only in memory. The hook key on disk can only ask, never answer.
Force-pushes, rm -rf and reading key files are denied unless you switch them off in Safety.
An "Always" shortcut never approves a high-risk request unless you say so. Critical requests can only be allowed once.
Honest note: risk badges are a warning, not a sandbox. AgentDeck only sees what an agent asks permission for, and a determined trick can get past a text check. Keep your agent's own permission settings sensible.
04Privacy
No telemetry. No analytics. No accounts. AgentDeck connects to nothing unless you switch it on.
The session files and logs of Claude Code and Codex, and the permission requests they send through their hooks.
Never their login credentials.
Your shortcuts, settings, usage numbers and history (kept 90 days), all in one folder:
%APPDATA%\AgentDeck
Nothing.
Unless you switch it on: update checks, CI status through your GitHub CLI, and your phone from anywhere, encrypted so the relay in between can't read it. All off by default.
This website too: no cookies, no analytics, no third-party requests.
05Setup
Seeing your sessions needs no setup at all. A short guide connects your agents with one switch each, and Settings tells you if anything needs a fix.
Double-click the installer, or install with one command. It installs for your user only, no admin rights, and starts with Windows (you can turn that off).
The beta isn't code-signed yet, so Windows may warn you.
More infoRun anyway
One switch each. AgentDeck saves a backup of your agents' settings first.
Hover the top-center edge of your screen to peek, click to open. It drops down by itself when an agent needs you.
06FAQ
Something else? Use Send feedback in the app.
Yes, it's free to use while in beta. It isn't open source; the license agreement is shown by the installer.
Claude Code and Codex, on Windows 10 and 11. Seeing sessions needs no setup; answering requests from the island needs the hooks the setup guide adds.
The beta isn't code-signed yet. Click More info → Run anyway. It installs for your user only, with no admin rights.
Nothing changes for your agents: they ask in their terminal as they always do. AgentDeck never approves anything on its own when something goes wrong, and a request nobody answers within 9 minutes goes back to the terminal.
No. There's no telemetry, no analytics and no account. The network is used only by features you switch on: update checks, CI status through your own GitHub CLI (gh), and your phone from anywhere, where everything is encrypted between your phone and your PC. The details.
Yes, with the Android app. In Settings → Phone, click Connect a phone and scan the code with your phone's camera: the page that opens installs the app and pairs it. On your Wi-Fi it talks to your PC directly. Turn on From anywhere to reach it away from home and get a notification when an agent needs you. It works on phones without Google services too. iPhone comes later; until then the page works in any phone's browser on your Wi-Fi.
Only to add its own hooks (and, if you want, a status line for Claude's usage meter), with a backup saved first. If you already have your own status line, AgentDeck leaves it alone. Uninstalling removes only AgentDeck's entries; your own settings are left exactly as they were.
Claude only shares usage through a status line, in terminal sessions, on Pro and Max plans. Turn on Show my Claude usage in Settings.
AgentDeck checks the format of everything it reads. If an update changes something, only that feature turns off, with a plain message. No guessed numbers.
Free while in beta. Windows 10 & 11.
AgentDeck_x64-setup.exe · v0.18.0 · 5 MB · installs per user, no admin rights
Release notes and older versions
Or install with one command in PowerShell: no browser warnings.
$f="$env:TEMP\AgentDeck-setup.exe"; iwr https://github.com/hicham-alaoui0/agentdeck-beta/releases/latest/download/AgentDeck_x64-setup.exe -OutFile $f -UseBasicParsing; start $f
Downloads the latest installer from this site's GitHub releases and opens it.
Why: the beta isn't code-signed yet and is new, so Windows doesn't know it. Updates from inside the app skip all of this.
AgentDeck is in a small private beta right now: the installer is shared by invitation.