RunanywhereAI/RCLI

What it solves

RCLI provides a fully local, on-device voice AI assistant for macOS that eliminates the need for cloud APIs and internet connectivity. It allows users to control their computer via voice, query their own documents, and analyze visual information from the camera or screen without sending data to external servers.

How it works

It implements a complete AI pipeline consisting of Speech-to-Text (STT), a Large Language Model (LLM), and Text-to-Speech (TTS), all running natively on Apple Silicon. The system uses a proprietary GPU inference engine called MetalRT (for M3 chips and later) or falls back to llama.cpp (for M1/M2 chips) to ensure low latency. It integrates with macOS via AppleScript and shell commands to execute system actions and uses a hybrid vector + BM25 retrieval system for local RAG (Retrieval-Augmented Generation) over PDFs, DOCX, and text files.

Who it’s for

Mac users with Apple Silicon (M1, M2, M3, M4) who want a private, low-latency voice assistant capable of system control, document intelligence, and on-device vision analysis.

Highlights

• Local Execution: 100% on-device inference with no cloud or API keys required.
• System Control: 40 built-in macOS actions for productivity, communication, media, and system settings.
• Multimodal Capabilities: On-device VLM (Vision-Language Model) for analyzing images, live camera feeds, and screen regions.
• High Performance: Sub-200ms end-to-end voice latency and high LLM throughput via the MetalRT engine.
• Document Intelligence: Local RAG for querying indexed documents with ~4ms retrieval latency.
• Interactive TUI: A terminal-based dashboard for managing models, voices, and system actions.