Radar for Claude Code

💬 working · ⚠️ needs you · 🟢 done · 🔴 failed · 🤖 agents

Claude Code multitasking, done right. Run Claude across several projects and always know which one needs you, without clicking through window after window.

Four features, one glance

• Live status in every window title. 💬 working, ⚠️ needs you, 🟢 done, 🔴 the turn failed. Every project, one look. A question Claude typed counts as ⚠️ too, so it doesn't sit unanswered behind a green marker.
• Live agent count. Claude fans out into subagents for a big review or a workflow? The title counts them while they run: 💬 14m 🤖8.
• Markers that clean up after themselves. Interrupted with Esc, Claude quit, terminal closed? The marker goes, instead of claiming work that stopped ten minutes ago. Hit a spend limit or a 529 in a background window? You get told, instead of watching 💬 count minutes in which nothing can happen.
• macOS banners and sounds. A background window pings you when it needs you or finishes, one click jumps right in. The window you're looking at stays quiet.

Tips

• Native tabs are the killer setup. Set "window.nativeTabs": true in your settings and restart: every window packs into one tab bar, a live status board for all your Claudes. Already have windows open? Window → Merge All Windows pulls them all into one bar. (Radar works without native tabs too: title bar, window switcher, Mission Control.)
• Switch projects without the mouse. Bind workbench.action.showPreviousWindowTab and showNextWindowTab to keys you like (⌘← / ⌘→ work nicely), and step straight to the Claude that needs you. Without native tabs, ⌘` cycles through the windows.
• Presenting? Run Mute banners for 1 hour from the Command Palette, sounds pause too and the markers keep working. For quiet by default, empty claudeRadar.soundWaiting / soundDone or turn claudeRadar.banner off.

Requirements

• VS Code ≥ 1.93 · Claude Code ≥ 2.1.247 · macOS only
• terminal-notifier (Homebrew) makes banners clickable; without it you still get notification + sound

Setup

Radar offers to set itself up on first launch and keeps reminding you until it's done: both steps are required, no marker shows without them. If terminal-notifier is missing, it offers the Homebrew install too. Or do it via the Command Palette: Add ${claudeRadarStatus} to window.title, then Install Claude hooks (writes to ~/.claude/settings.json, backup first, other hooks stay put). Claude loads hooks at session start, so restart any session that's already running. Updates keep the hooks current automatically. Uninstalling? Run Remove Claude hooks first, then uninstall.

Privacy, performance & trust

• No network, no telemetry. Everything happens in local files.
• Event-driven, not polling. Idle cost is effectively zero. One exception: while a marker is on display, a light scan runs every 10 s (agent count, and whether a Claude is still alive here). With an empty title, the normal state, nothing runs at all.
• No latency for Claude. The hook after every tool call runs async, so Claude never waits for Radar; the other hooks fire once per turn.
• Safe config writes. Atomic, with a backup. Removal is one command.
• Open source. Full code on GitHub.

How it works

Eight Claude hooks drop a file in ~/.claude/tab-status/, named sha256(project path) (first 16 hex chars, realpath-resolved so symlinks match). The content is the status:

• UserPromptSubmit → prompt (shows working, purely visual, no banner/sound; its own label so a new prompt restarts the elapsed-time clock)
• Notification / PermissionRequest → waiting
• PreToolUse + AskUserQuestion → waiting (a question needs you too; its own hook, since AskUserQuestion fires neither Stop nor Notification)
• PostToolUse / PostToolUseFailure → working (every completed tool call flips back to working, failed ones too, since Claude reads the error and carries on; so an answered question or a granted permission doesn't leave the title empty while Claude keeps going)
• Stop → stop (done, or ⚠️ when the turn ends by asking you something, or 🔴 when its last tool call failed), unless background work is still running, then working stays. Subagents, workflows, and monitors count as work (they finish and wake the session); background shell tasks don't, since a dev server Claude started would otherwise hold the working marker forever. Want shell tasks to count? Turn on claudeRadar.shellTasksKeepWorking.
• StopFailure → error (🔴, the turn died on Claude's side: a spend limit, a 529, an expired session, a connection lost mid-response. The banner names the cause, so you know whether to retry or to go fix something.)

Each window watches only its own file and shows the marker from state + focus:

• working stays visible the whole time, even in the active window, never with banner/sound, so nothing flickers.
• needs you / done show a persistent marker and banner only when the window's in the background (banner via terminal-notifier, a click focuses the right window). In the active window they flash for ~1 s ("peek") and fade.
• Focusing a window clears an attention marker; working stays, since Claude's still going.

While working shows, Radar also counts the subagents running in this project: Claude Code keeps one transcript file per agent under ~/.claude/projects/, and an agent counts as running until its transcript ends on its final answer (checked every 10 s, read-only). One that has been silent for ten minutes counts as gone, whatever its transcript says. That's the 🤖8 in the title.

While a marker is on display, Radar also verifies it. Claude Code lists every running session in ~/.claude/sessions/<pid>.json with its working directory; if no live process belongs to this project any more, the marker was lying and goes. The same pass reads the tail of the session transcript, which records two things no hook reports:

• An Esc interrupt. The marker goes, deliberately without a "done" marker, since nothing finished.
• A failed tool call. Only the turn's last tool result counts for 🔴, so a failure Claude recovered from in the same turn still ends green. Measured across 1464 real turns, that marks 2 % of them.

Since Claude Code 2.1.269 the same registry entry also carries the session's own status, and the sweep reads that too. Every session of the project idle since your prompt means the turn ended without a Stop, and the marker goes. A session waiting for a permission or a dialog shows ⚠️, banner included, whether or not a hook reported it. This is a backstop for the hooks, not a replacement: a hook reports the moment it happens, the sweep within 10 s.

Stale files (sessions with no matching window) are cleaned up after 24 h.

Limits

• The marker is text (emoji) in the title, not a native tab badge, macOS doesn't allow those.
• Stop fires after every response. Fine for the marker and banner (idempotent, one banner per project, focused windows skipped).
• Interrupting Claude with Esc still fires no hook Radar can use, so the marker clears on the next sweep (within 10 s) rather than instantly.
• Marker verification needs ~/.claude/sessions/, which older Claude Code versions don't write, and the status backstop needs the status field in it (Claude Code 2.1.269). Without them Radar leaves every marker alone, exactly as before.
• 🔴 means the turn failed: either its last tool call failed, or Claude's side died on it. A task Claude explains it can't do ends green, since nothing technically failed.
• Question detection reads the last non-empty line of Claude's closing text. A question phrased mid-paragraph, or one without a question mark, ends green. Measured across 7528 real turn endings, 3 % close on a question.
• The agent count reads the transcripts: an agent killed together with its session leaves no final answer behind and stays in the count for up to ten minutes. Agents running in Anthropic's cloud leave no local transcripts and aren't counted.
• Multi-root workspaces: the first folder wins.
• Two events at once: the later one wins.

Settings

Give just the symbol, the space is added for you. Empty means no marker for that status.