Wire a desktop pet into what your Claude Code session is actually doing — all 15 hook events.

中文 · macOS · MIT

_{When Claude Code asks you to pick between options, the card lands on the desktop — and the pet switches to its lightbulb "needs input" pose.
The card follows Clawd's own language setting; English and Chinese both shown}

Clawd on Desk is a desktop pet that watches AI coding agents. It ships with Claude Code integration and works out of the box.

This repo is a complete wiring layer on top of it: all 15 Claude Code lifecycle hook events, plus a notification script, an idempotent install/uninstall toolchain, and the parameter trade-offs I settled on after three months of daily use.

The difference is granularity. With fewer events the pet roughly knows whether you're busy or idle. With all 15 it can tell you're running a tool, that a tool just failed, that subagents are running in parallel, that context is being compacted — or that it's blocked on a confirmation waiting for you.

That last one is the whole point.

During long runs you leave the terminal — docs, messages, coffee. You come back and find it stopped three minutes ago waiting for you to click "allow."

This config cuts those three minutes to zero:

• Permission requests surface as a desktop card: Cmd+Shift+Y to allow, Cmd+Shift+N to deny — no window switching
• The card never auto-dismisses (permissionBubbleAutoCloseSeconds: 0). Auto-dismiss is the worst default here: you think you denied it, but it just timed out and fell back to the terminal prompt
• Dock flash + sound on completion, so you catch it from another app
• Tool failures change the pet's state immediately — no scrolling back to find which step blew up

_{When the command contains rm -rf, the card raises a Destructive action warning on its own}

Claude Code fires hooks at key points in a session's lifecycle. This config forwards 14 of them as command hooks to Clawd's bundled clawd-hook.js, and routes one more — the permission request — as an http hook straight to Clawd's local port.

Claude Code                          Clawd on Desk
    │
    ├─ SessionStart ─┐
    ├─ PreToolUse  ──┤  command hook
    ├─ Stop        ──┼─→ clawd-hook.js ──→ POST 127.0.0.1:23333/state ──→ pet changes state
    ├─ … (14 total) ─┘        (async, 5s timeout, never blocks the session)
    │
    └─ PermissionRequest ─→ HTTP hook ──→ POST 127.0.0.1:23333/permission
                                              (blocking, 600s timeout)
                                                    │
                                          card pops ┴─→ you click Allow / Deny
                                                         └─→ decision returns, session continues

The two channels differ in whether they wait for you. State events are one-way broadcasts — fire and forget. A permission request has to stop and wait for a human decision, which is why it is the only blocking one.

Taken from Clawd's clawd-hook.js (EVENT_TO_STATE) — not guesswork:

_{The four most distinguishable states. sweeping means context is being compacted, attention means the turn ended and it came to get you, error literally smokes.
The rest (thinking/working/juggling) differ mostly in motion and are hard to tell apart in a still frame}

_{Left: sweeping, during context compaction. Right: error, after a tool failure.
error is transient — it reverts to working after roughly a second, so you have to be looking to catch it}

One detail worth knowing: on some builds Claude Code reports subagent launches only as PreToolUse(Task) without a native SubagentStart. Clawd handles this by switching to juggling on the Task / Agent tool name — so parallel work never goes unreported.

"command": "\"/opt/homebrew/bin/node\" \"/Applications/Clawd on Desk.app/…/clawd-hook.js\" Stop"

Hooks run in a non-login shell, where PATH may not include Homebrew. A bare node becomes command-not-found — and because these are async: true, you never see the error. The symptom is just a pet that quietly stops moving. install.sh detects and fills both paths for you.

The path contains a space (Clawd on Desk.app), so every segment must be quoted.

# 1. Install Clawd on Desk itself (not bundled or redistributed here)
#    https://github.com/rullerzhou-afk/clawd-on-desk/releases

# 2. Apply this config
git clone https://github.com/ssssssanjiu/clawd-conduit.git
cd clawd-conduit
bash install.sh

Start a new Claude Code session to pick it up.

1. Locate Clawd — checks /Applications, then ~/Applications, then falls back to mdfind on bundle id com.clawd.on-desk. Fails loudly with a download link if it can't find it.
2. Locate node — checks /opt/homebrew/bin/node, /usr/local/bin/node, then command -v node.
3. Back up — copies your settings.json to settings.json.bak.<timestamp>.
4. Merge hook config — fills the placeholder template with real paths. Preserves your other hooks per event, replacing only entries this repo wrote itself.
5. Install the notification script — copies it to ~/.claude/hooks/ and attaches it to Notification (skippable).
6. Validate — runs python3 -c "json.load(...)" to confirm the JSON it wrote is still parseable.

Safe to re-run. A second run replaces rather than appends; the occurrence count of clawd-hook.js stays at exactly 14.

# 1. Are all 15 events wired?
python3 -c "import json;h=json.load(open('$HOME/.claude/settings.json'))['hooks'];\
print(len([k for k,v in h.items() if 'clawd' in json.dumps(v).lower() or '23333' in json.dumps(v)]),'events')"

# 2. Is Clawd's local service listening?
lsof -nP -iTCP:23333 -sTCP:LISTEN

# 3. Does the hook script path actually exist?
ls -l "/Applications/Clawd on Desk.app/Contents/Resources/app.asar.unpacked/hooks/clawd-hook.js"

The first should print 15 events; the second should show a Clawd on Desk process. Once both check out, start a session and send any message — the pet should go from idle to thinking.

ls -t ~/.claude/settings.json.bak.* | head -1   # most recent

Keep manageClaudeHooksAutomatically enabled in Clawd's settings — it repairs the app paths inside your hooks after a Clawd upgrade, so you don't have to re-run install.sh every time.

├── install.sh                        # path detection + idempotent merge
├── uninstall.sh                      # clean removal, keeps your other hooks
├── settings/
│   └── clawd-hooks.template.json     # all 15 events, paths as placeholders
├── hooks/
│   └── notify-input-needed.py        # macOS banner + Glass sound
└── docs/
    ├── hook-events.md                # every event explained, and why absolute paths
    └── recommended-prefs.md          # the Clawd settings I actually run, with reasons

hooks/notify-input-needed.py, about 20 lines:

On a Claude Code Notification event it reads the event JSON from stdin, pulls out the message, and fires a system banner with the Glass sound via osascript.

It complements Clawd's own bubble rather than replacing it: Clawd's bubble is on-screen and interactive; the system banner lands in Notification Center and reaches you inside fullscreen apps. Run both and you miss the least.

The reason it's needed: Clawd's "passive notification" bubbles serve only agents without a decision channel, like Codex and Kimi (see passive-notify-entry.js). Claude Code goes through the decision-carrying path, so the only cards you ever see are permission cards and follow-up cards — pure state changes produce no bubble at all. This script fills that gap.

Skip it with INSTALL_NOTIFY=0 bash install.sh.

bash uninstall.sh

Removes the hook config only; leaves the Clawd app untouched. Backs up first. Round-trip tested: after uninstalling, settings.json is byte-for-byte identical to what it was before installing.

The pet itself is Clawd on Desk (@rullerzhou-afk, AGPL-3.0-only). You need it installed before this config does anything. If you like it, go star the upstream repo.

This repository bundles and redistributes none of Clawd's code or binaries — only config templates, install scripts, and docs.

MIT. See LICENSE and NOTICE.md.