LokalBot is a new macOS application that performs meeting recording and summarization, typing autocomplete, and daily activity tracking entirely on-device without requiring cloud services or API keys. The tool is designed specifically for Apple Silicon Macs running macOS 15 or later, leveraging the Neural Engine, MLX, and Core Audio to ensure all processing remains offline.
- The app uses llama.cpp with libllama for low-latency cotyping and llama-server for other tasks, supporting GGUF files, Ollama, or OpenAI-compatible endpoints.
- Transcription is handled by Granite Speech 4.1, Parakeet, Whisper, or Qwen3-ASR via CoreML/MLX, with Parakeet achieving approximately 190x realtime speed.
- Semantic search utilizes a Qwen3-Embedding 0.6B GGUF model on a separate llama-server instance, storing vectors in SQLite for instant local retrieval.
- Optional diarization is available via pyannote through FluidAudio to distinguish between different speakers.
- Privacy is prioritized by eliminating network calls after the initial model download, with optional screenshots sealed via AES-GCM and set to auto-delete.
The application aims to consolidate multiple separate tools into a single interface while maintaining strict privacy by keeping all data on the user's machine.