Changelog

Notable changes to Duckwork, newest first.

0.3.0 13 September 2026

• Update Info.plist
• Remove editable Full Assistant prompt
• Ship in-process Qwen3.5-4B local backend
• Shift speech stack to Whisper and local TTS
• Add on-device Qwen3-TTS via TTSKit and model download UX
• Simplify speech and history settings
• Remove diagnostics clipboard fallback
• Keep onboarding windows visible and model-ready
• Switch transcription to WhisperKit/Core ML

0.2.0 12 September 2026

• Use SQLite (FTS5) for session storage
• Per-line transcript corrections with history
• Remove daily digest feature
• Add session memory and MCP serving
• Unify live transcript handling
• Track partial speech timestamps
• Update CI and fix Swift concurrency issues
• Preserve utterance timestamps and mark retroactive
• Harden core session flow and ship 0.2.0
• Avoid repeated Echo on unchanged transcript
• Make Settings a resizable window with icon tabs
• Cap response text height in menu popover
• Clarify escalation and mode behavior
• Show MCP tool count under Settings
• Expose turn silence and add mode description

0.1.0 6 September 2026

First public release. Duckwork is a macOS menu bar app for thinking out loud: it listens on a global hotkey and stays quiet until you ask for a response.

Capture

• Start and stop listening from anywhere with a global hotkey.
• On-device speech recognition by default, using Apple's Speech framework — audio never leaves your Mac.
• Optional server-based recognition for higher accuracy.
• Session transcripts stored locally, with a browsable history and per-session summaries. Delete any session from within the app.

Modes

• Silent — the default. Nothing is sent anywhere and nothing answers. Pauses stay silent.
• Echo — paraphrases what you just said in a sentence or two. Triggered with ⌥⌘E or a pause.
• Full Assistant — a full answer with optional MCP tool use, spoken aloud if you want. Triggered with ⌥⌘F or a pause.
• Escalate from any mode with a hotkey for a single response, then drop straight back to where you were.

Models and voices

• Bring your own backend: Ollama for fully local models, any OpenAI-compatible endpoint, or the Anthropic API directly.
• Switch provider at any time; API keys are stored in the macOS Keychain.
• Speak responses with any voice installed on your Mac.
• Connect MCP servers so Full Assistant mode can act on what you are working on.