RCLI Brings a Full Voice Assistant to Your Mac Without any Cloud
RCLI packs a full speech-to-speech pipeline plus RAG and vision into a single MIT-licensed Mac binary that runs entirely offline.
Takeaways
- RCLI is a MIT-licensed on-device voice AI CLI for macOS with STT, LLM, TTS, VLM, and RAG. GitHub
- Powered by MetalRT, a hand-written Metal GPU engine reporting 658 tok/s decode on M4 Max. Benchmarks
- Delivers sub-200ms end-to-end voice latency and claims 714x faster-than-real-time Whisper transcription.
- Ships 40 macOS voice actions covering Spotify, notes, reminders, dark mode, and web search via tool calling.
- Local RAG over PDFs and DOCX with hybrid vector plus BM25 retrieval at roughly 4ms per query.
- Requires Apple Silicon M3+ for fast path; M1 and M2 fall back to llama.cpp automatically.
A new open-source project called RCLI aims to prove that a full voice assistant, complete with document Q&A and computer control, can live entirely on your laptop. It ships as a terminal app for macOS that wires together speech-to-text, an LLM, text-to-speech, and a vision model into one pipeline, with no cloud calls and no API keys. The repo has crossed 1.5k stars and sits on top of a proprietary Metal inference engine that the authors claim beats both llama.cpp and Apple's own MLX on Apple Silicon.
What it actually does
RCLI is a command-line tool that turns your Mac into a local voice assistant. Hit space in the TUI, talk, and it transcribes, reasons, replies aloud, and can execute macOS actions on your behalf. The full feature set covers four modalities running natively on Metal:
- Voice pipeline with Silero VAD, Zipformer or Whisper or Parakeet for STT, Qwen3 or LFM2 for the LLM, and Kokoro or Piper for TTS running on three concurrent threads.
- Vision through
rcli vlm photo.jpg, live camera capture, or screen-region analysis using Qwen3 VL 2B, LFM2 VL 1.6B, or SmolVLM 500M. - 40 macOS actions exposed as LLM tools, covering Spotify playback, volume, dark mode, screenshots, notes, reminders, FaceTime calls, and web searches via AppleScript and shell.