Arcade Agents watches every AI coding agent you run, shows you what each one is doing, and lets you approve — or stop — what it wants, without leaving your editor.
Claude Code · Codex · Grok Build · Kiro CLI
The real thing, at real size — click a state, try Allow or Deny, and watch a finished agent take its lap.
From the moment an agent starts to the moment it needs you — all in a strip of black the width of your notch.
Not a log of what finished — what's happening now. The notch reads Running Bash while the command is still running, so a long build looks busy instead of looking dead.
Claude, Codex, Grok and Kiro each get their own body — you can tell them apart at a glance. The shape says which tool it's using; the colour says whether you're needed: pink working, green done, yellow waiting on you.
When an agent wants to run a command or edit a file, the island opens with the command and two buttons. Allow or deny without touching your terminal.
Every request is scored by blast radius. rm -rf and force-pushes light up red with the reason spelled out; a git status barely registers. Tick "always allow" on the boring ones and it stops asking — but a destructive command can never become a rule.
The gate already waits up to 24 hours for a human — so answering from the kitchen costs nothing. Scan a QR, and pending approvals show up on your phone with the same risk read. Local network only. No cloud, no account, off by default.
Every tool call and every decision, written to disk as it happens. Come back from lunch and read exactly what ran — or hand the CSV to whoever asks whether agents can be trusted against production.
Tools run, approvals granted, and a trust score — the share of an agent's requests you actually said yes to. An agent you keep denying is one you don't trust, and now you can see it.
The safety machinery is real. The way it tells you about it isn't po-faced.
When an agent completes, its creature goes green and runs the length of the island — hopping, throwing off pixel sparks. A second and a half, then it's back to being a quiet ear. It's how you catch a finished run out of the corner of your eye instead of alt-tabbing to check.
When something scores critical — a force push, a recursive delete — the island doesn't quietly ask. The card turns into a boss fight, the creature grows, and the reason is spelled out in plain English. Denying it is the win, and the scoreboard counts it.
Run Claude and Grok in the same project and they will happily stomp each other's edits — neither one knows the other exists. Arcade Agents does, and it says so, loudly, before it costs you an hour. It also tracks your git worktrees, so you can give each agent its own.
It writes one hook into each agent's own config — your existing hooks preserved, every change backed up first.
$4.99 / year · cancel anytime · no account, nothing leaves your Mac.
Claude Code, Codex and Grok Build today, with real approvals and live status — Kiro CLI is on the way. Each one speaks a different hook dialect, and Arcade Agents translates all of them, so an approval genuinely blocks the agent until you decide.
Because it has to. Grok exposes only one place to intercept a tool call, so every call passes through the gate — where Claude and Codex have a dedicated approval event and everything else flows straight through. Set Auto-approve read-only in Settings and the boring ones stop asking; teach it a rule and they stop for good.
No. There's no cloud, no account, and no telemetry. Agents talk to the app over a local socket. Phone approval, when you turn it on, serves a page on your own Wi-Fi behind a fresh token — it never touches the internet.
When an agent wants to run a tool, its hook calls a tiny bridge that opens the island and waits. You click Allow or Deny, and the agent — blocked the whole time — proceeds or stops. If the app isn't running, the bridge gets out of the way instantly so nothing ever hangs.
It falls back to a floating pill at the top-center of your screen, and rides above fullscreen apps so it's there even when you're heads-down in an editor.
No. Native Swift, no Electron, under 50MB of RAM and near-zero CPU when idle. The characters and the 8-bit sounds are both generated in code — there isn't a single image or audio file in the app.
That's the point. Every session shows up as its own creature in the notch, approvals queue so none get lost, and the Worktrees view tells you which agent is live in which git worktree. If two agents end up in the same project, it warns you — they can't see each other, but it can.
The app is signed with a Developer ID certificate. The first time you open it, Control-click the app in Applications and choose Open — after that it launches normally. It's a one-time step.