What it does Ghost Pepper is a menu-bar macOS app that transcribes your voice and meetings using open-source Whisper and Qwen models running locally on Apple Silicon. Hold Control, speak, release; it pastes the cleaned-up text wherever your cursor is. For meetings, it records, transcribes, and spits out a markdown summary — all without touching a cloud API.

The interesting bit The author didn’t just claim “local-only.” He had Claude review the entire codebase, published the exact audit prompt and file-level results in PRIVACY_AUDIT.md, and dares you to rerun it yourself. That’s a refreshingly paranoid posture in a space where competitors have raised $80M to do the same thing with servers.

Key highlights

• Zero-cloud by default: Core features use WhisperKit, LLM.swift, and Apple Vision; optional cloud integrations (Zo AI, Trello) are disabled and require your own keys.
• Model buffet: Five speech models from 75 MB English-only Whisper tiny up to 900 MB Qwen3-ASR for 50+ languages; three cleanup LLMs sized for speed vs. quality tradeoffs.
• Enterprise-friendly: Ships with exact Bundle ID, Team ID, and PPPC payload specs for MDM pre-approval — a nice touch for managed Macs.
• Open source, MIT licensed: Build from GhostPepper.xcodeproj in Xcode if you don’t trust the DMG.

Caveats

• Apple Silicon only, macOS 14+: Intel Macs need not apply.
• Accessibility permission required: The global hotkey and simulated keystrokes need it, which means Gatekeeper warnings and potential IT friction on locked-down machines.
• Model downloads are one-time but from Hugging Face: Not a privacy leak, but your first run pulls ~500 MB–1.4 GB depending on model choice.

Verdict If you want dictation or meeting notes without wondering where your voice data sleeps, this is a strong contender — especially since the code is auditable and the price is zero. If you’re on Intel, need cross-platform sync, or want polished SaaS polish, look elsewhere.