Duck Duck Duck
A companion for Claude Code on Mac. It watches your coding sessions, scores every prompt and response, speaks opinionated reactions, and handles permissions by voice. Optionally connects to a physical duck for hardware reactions.
Learn more at duck-duck-duck.edges.ideo.com
🔒 Default intelligence is fully on-device and private on M3+ Macs. M1/M2 users get the best experience with a cloud API key (free options available). No cloud audio regardless — your voice never leaves your machine.
Requirements
- macOS 26 (Tahoe) or later, Apple Silicon (M1 or later)
- Claude Code or Claude Desktop — get Claude Code
On-device scoring is designed for M3+ Apple Silicon. It works on M1/M2 but runs slowly (~30-60s per eval). For instant reactions on older hardware, add a cloud API key — see Evaluation below.
Install
- Download
DuckDuckDuck.dmg from GitHub Releases
- Drag to Applications, launch
- Grant Microphone and Speech Recognition when prompted — all audio stays on-device
- Accept the Terms of Use
- The app walks you through installing Claude and the plugin
- Open a Claude Code session (CLI or Desktop) — the duck is watching
Build from source
git clone https://github.com/ideo/Rubber-Duck.git
cd Rubber-Duck/widget
make run
Then use Setup → Install Plugin from the menu bar.
Requires Xcode with Swift 6.2+ (macOS 26 SDK).
How It Works
You ──► 🦆 Hardware Duck ──► Duck Widget (SwiftUI) ◄── Claude Code / Desktop
mic + speaker eval engine (Foundation plugin hooks (prompt,
servo + LED Models / Haiku / Gemini) response, permission)
▲ │
│ voice out (TTS)
└────────────────── servo + LED commands
speaker audio
(no Duck, Duck, Duck device? laptop mic + speakers work too)
- Hooks fire on Claude Code events and POST to the widget's embedded server
- Eval engine scores text on-device via Apple Foundation Models (free, sub-second on M3+) — returns scores + a spoken reaction
- Widget animates the duck face, speaks the reaction, optionally drives hardware via USB serial
- Voice permissions — the duck summarizes what Claude wants to do and asks. Say "yes", "always allow", "deny". Foundation Models classifies ambiguous responses.
- Voice commands — say "ducky [command]" to inject text into Claude Code via tmux
- Wildcard voice — score-gated AI picks from 10 voices per reaction (normal, grave, cheerful, dramatic, whisper, etc.)
Modes
| Mode | What it does | Mic |
|---|
| Companion | Reacts to everything, voice permissions, wake word | On |
| Permissions Only | Silent until a permission arrives | On |
| Companion (No Mic) | Reacts and speaks, click-only permissions | Off |
| Relay (Experimental) | Speak directly to Claude CLI via tmux | On |
Evaluation
Each prompt and response is scored from -1.0 to +1.0:
| Dimension | What it measures |
|---|
| creativity | Novel/surprising vs boring/obvious |
| soundness | Technically solid vs flawed |
| ambition | Bold undertaking vs trivial tweak |
| elegance | Clean/clear vs hacky/convoluted |
| risk | Could break things vs safe |
Defaults to Apple Foundation Models (on-device, free, designed for M3+). Switch to Claude Haiku or Gemini Flash from the menu bar for higher-quality scoring. See Data & Privacy for details.
M1/M2 Mac? Use a cloud API key for snappy reactions
On-device scoring runs slowly on M1/M2 (~30-60 seconds per eval). For instant results:
Gemini Flash (free tier, no credit card):
- Go to aistudio.google.com and sign in with Google
- Click Get API Key → Create API Key
- Paste it into Preferences → Intelligence → Gemini
Claude Haiku (~$0.001 per eval):
- Go to console.anthropic.com and create an account
- Add a payment method (pay-as-you-go)
- Go to API Keys → Create Key
- Paste it into Preferences → Intelligence → Anthropic
Voice
