Talk to your Mac, query your docs, no cloud required.
RCLI is an on-device voice AI for macOS. A complete STT + LLM + TTS pipeline running natively on Apple Silicon — 43 macOS actions via voice, local RAG over your documents, sub-200ms end-to-end latency. No cloud, no API keys.
Powered by MetalRT, a proprietary GPU inference engine built by RunAnywhere, Inc. specifically for Apple Silicon.
Demo
Real-time screen recordings on Apple Silicon — no cloud, no edits, no tricks.
|
Voice Conversation Talk naturally — RCLI listens, understands, and responds on-device. 
Click for full video with audio |
App Control Control Spotify, adjust volume — 43 macOS actions by voice. 
Click for full video with audio |
|
Models & Benchmarks Browse models, hot-swap LLMs, run benchmarks — all from the TUI. 
Click for full video with audio |
Document Intelligence (RAG) Ingest docs, ask questions by voice — ~4ms hybrid retrieval. 
Click for full video with audio |
Install
Requires macOS 13+ on Apple Silicon (M1 or later).
One command:
curl -fsSL https://raw.githubusercontent.com/RunanywhereAI/RCLI/main/install.sh | bash
Or via Homebrew:
brew tap RunanywhereAI/rcli https://github.com/RunanywhereAI/RCLI.git
brew install rcli
rcli setup
Quick Start
rcli # interactive TUI (push-to-talk + text)
rcli listen # continuous voice mode
rcli ask "open Safari" # one-shot command
rcli ask "play some jazz on Spotify"
Features
Voice Pipeline
A full STT + LLM + TTS pipeline running on Metal GPU with three concurrent threads:
- VAD — Silero voice activity detection
- STT — Zipformer streaming + Whisper / Parakeet offline
- LLM — Qwen3 / LFM2 / Qwen3.5 with KV cache continuation and Flash Attention
- TTS — Double-buffered sentence-level synthesis (next sentence renders while current plays)
- Tool Calling — LLM-native tool call formats (Qwen3, LFM2, etc.)
- Multi-turn Memory — Sliding window conversation history with token-budget trimming
43 macOS Actions
Control your Mac by voice or text. The LLM routes intent to actions executed locally via AppleScript and shell commands.
| Category | Examples |
|---|---|
| Productivity | create_note, create_reminder, run_shortcut |
| Communication | send_message, facetime_call |
| Media | play_on_spotify, play_apple_music, play_pause, next_track, set_music_volume |
| System | open_app, quit_app, set_volume, toggle_dark_mode, screenshot, lock_screen |
| Web | search_web, search_youtube, open_url, open_maps |
Run rcli actions to see all 43, or toggle them on/off in the TUI Actions panel.
Tip: If tool calling feels unreliable, press X in the TUI to clear the conversation and reset context. With small LLMs, accumulated context can degrade tool-calling accuracy — a fresh context often fixes it.
RAG (Local Document Q&A)
Index local documents, query them by voice. Hybrid vector + BM25 retrieval with ~4ms latency over 5K+ chunks. Supports PDF, DOCX, and plain text.
rcli rag ingest ~/Documents/notes
rcli ask --rag ~/Library/RCLI/index "summarize the project plan"
Interactive TUI
A terminal dashboard with push-to-talk, live hardware monitoring, model management, and an actions browser.
| Key | Action |
|---|---|
| SPACE | Push-to-talk |
| M | Models — browse, download, hot-swap LLM/STT/TTS |
| A | Actions — browse, enable/disable macOS actions |
| B | Benchmarks — run STT, LLM, TTS, E2E benchmarks |
| R | RAG — ingest documents |
| X | Clear conversation and reset context |
| T | Toggle tool call trace |
| ESC | Stop / close / quit |
MetalRT GPU Engine
MetalRT is a high-performance GPU inference engine built by RunAnywhere, Inc. specifically for Apple Silicon. It delivers the fastest on-device inference for LLM, STT, and TTS — up to 550 tok/s LLM throughput and sub-200ms end-to-end voice latency.
Apple M3 or later required. MetalRT uses Metal 3.1 GPU features available on M3, M3 Pro, M3 Max, M4, and later chips. M1/M2 support is coming soon. On M1/M2, RCLI automatically falls back to the open-source llama.cpp engine.
MetalRT is automatically installed during rcli setup (choose "MetalRT" or "Both"). Or install separately:
rcli metalrt install
rcli metalrt status
Supported models: Qwen3 0.6B, Qwen3 4B, Llama 3.2 3B, LFM2.5 1.2B (LLM) · Whisper Tiny/Small/Medium (STT) · Kokoro 82M with 28 voices (TTS)
MetalRT is distributed under a proprietary license. For licensing inquiries: founder@runanywhere.ai
Supported Models
RCLI supports 20+ models across LLM, STT, TTS, VAD, and embeddings. All run locally on Apple Silicon. Use rcli models to browse, download, or switch.
LLM: LFM2 1.2B (default), LFM2 350M, LFM2.5 1.2B, LFM2 2.6B, Qwen3 0.6B, Qwen3.5 0.8B/2B/4B, Qwen3 4B
STT: Zipformer (streaming), Whisper base.en (offline, default), Parakeet TDT 0.6B (~1.9% WER)
TTS: Piper Lessac/Amy, KittenTTS Nano, Matcha LJSpeech, Kokoro English/Multi-lang
Default install (rcli setup): ~1GB — LFM2 1.2B + Whisper + Piper + Silero VAD + Snowflake embeddings.
rcli models # interactive model management
rcli upgrade-llm # guided LLM upgrade
rcli voices # browse and switch TTS voices
rcli cleanup # remove unused models
Architecture
Mic → VAD → STT → [RAG] → LLM → TTS → Speaker
|
Tool Calling → 43 macOS Actions
Three dedicated threads in live mode, synchronized via condition variables:
| Thread | Role |
|---|---|
| STT | Captures audio, runs VAD, detects speech endpoints |
| LLM | Generates tokens, dispatches tool calls |
| TTS | Double-buffered sentence-level synthesis and playback |
Key design decisions:
- 64 MB pre-allocated memory pool — zero runtime malloc during inference
- Lock-free ring buffers for zero-copy audio transfer
- System prompt KV caching across queries
- Hardware profiling at startup for optimal config
- Token-budget conversation trimming
- Live model hot-swap without restarting
src/
engines/ STT, LLM, TTS, VAD, embedding engine wrappers
pipeline/ Orchestrator, sentence detector, text sanitizer
rag/ Vector index, BM25, hybrid retriever
core/ Types, ring buffer, memory pool, hardware profiler
audio/ CoreAudio mic/speaker I/O
tools/ Tool calling engine with JSON schema definitions
actions/ 43 macOS action implementations
api/ C API (rcli_api.h)
cli/ TUI dashboard (FTXUI), CLI commands
models/ Model registries with on-demand download
Build from Source
CPU-only build using llama.cpp + sherpa-onnx (no MetalRT):
git clone https://github.com/RunanywhereAI/RCLI.git && cd RCLI
bash scripts/setup.sh
bash scripts/download_models.sh
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
cmake --build . -j$(sysctl -n hw.ncpu)
./rcli
All dependencies are vendored or CMake-fetched. Requires CMake 3.15+ and Apple Clang (C++17).
CLI Reference
rcli Interactive TUI (push-to-talk + text + trace)
rcli listen Continuous voice mode
rcli ask <text> One-shot text command
rcli actions [name] List actions or show detail
rcli rag ingest <dir> Index documents for RAG
rcli rag query <text> Query indexed documents
rcli models [llm|stt|tts] Manage AI models
rcli voices Manage TTS voices
rcli bench [--suite ...] Run benchmarks
rcli setup Download default models
rcli info Show engine and model info
Options:
--models <dir> Models directory (default: ~/Library/RCLI/models)
--rag <index> Load RAG index for document-grounded answers
--gpu-layers <n> GPU layers for LLM (default: 99 = all)
--ctx-size <n> LLM context size (default: 4096)
--no-speak Text output only (no TTS)
--verbose, -v Debug logs
Contributing
Contributions welcome. See CONTRIBUTING.md for build instructions and how to add new actions, models, or voices.
License
RCLI is open source under the MIT License.
MetalRT is proprietary software by RunAnywhere, Inc., distributed under a separate license.
Built by RunAnywhere, Inc.