RCLI - On-Device Voice AI for macOS

TL;DR

TL;DR: RCLI is an open-source on-device voice AI for macOS that runs a full STT + LLM + TTS + VLM pipeline on Apple Silicon — no cloud, no API keys, sub-200ms latency.

Source and Accuracy Notes

• Project page: runanywhere.ai ← visited and verified
• Source repository: github.com/RunanywhereAI/rcli ← read README end-to-end
• License: MIT (verified via LICENSE file on GitHub)
• Latest release: v0.3.7 (2026-03-15, verified via GitHub Releases API)
• Stars: 1,530 (as of 2026-07-23, GitHub API)
• HN launch thread: not found via Algolia search (tool may not have had a dedicated Show HN post)

What Is RCLI?

RCLI is a command-line voice AI tool for macOS. It runs a complete STT + LLM + TTS + VLM pipeline locally on Apple Silicon — no cloud dependency and no API keys required.

The project is built by RunAnywhere, Inc. and uses MetalRT, a proprietary GPU inference engine the company built specifically for Apple Silicon. MetalRT is closed-source and distributed under a separate license; the rest of RCLI is MIT-licensed open source.

On M1/M2 Macs, RCLI falls back to llama.cpp automatically.

Setup Workflow

Step 1: Install

One-command install:

curl -fsSL https://raw.githubusercontent.com/RunanywhereAI/RCLI/main/install.sh | bash
```bash

**Or via Homebrew:**

```bash
brew tap RunanywhereAI/rcli https://github.com/RunanywhereAI/RCLI.git
brew install rcli
```bash

### Step 2: Download models

```bash
rcli setup
```bash

This downloads the default model bundle — approximately 1GB total, a one-time download. The bundle includes:

- LFM2 1.2B (default LLM)
- Whisper base.en (speech-to-text)
- Piper Lessac (text-to-speech)
- Silero VAD (voice activity detection)
- Snowflake embeddings

VLM models are downloaded on demand via `rcli models vlm`.

### Step 3: Run

```bash
rcli
```bash

This opens the interactive TUI in push-to-talk mode. Press to record, release to send.

## MetalRT GPU Engine

MetalRT is the custom GPU inference engine that sets RCLI apart from a plain llama.cpp setup. It is built by RunAnywhere specifically for Apple Silicon and accelerates the LLM inference path.

**Requirements:**
- macOS 13+
- Apple Silicon (M3 or later for MetalRT)
- M1/M2 Macs fall back to llama.cpp automatically

MetalRT is proprietary and closed-source. It is not included in the open-source build from source path.

## Local RAG Over Documents

RCLI includes a built-in RAG pipeline for querying your documents by voice:

```bash
# Index a directory of documents
rcli rag ingest ./my-docs/

# Query indexed documents
rcli rag query "what was discussed in last week's meeting?"
```bash

The retrieval step runs at approximately 4ms per query (hybrid retrieval, per the README). The LLM answer is generated locally.

## Supported Models

RCLI supports 20+ models across multiple categories:

| Category | Models |
|---|---|
| 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) |
| TTS | Piper Lessac/Amy, KittenTTS Nano, Matcha LJSpeech, Kokoro English/Multi |
| VLM | Qwen3 VL 2B, Liquid LFM2 VL 1.6B, SmolVLM 500M |

Model management commands:

```bash
rcli models          # interactive model browser
rcli models vlm      # download/manage vision-language models
rcli voices          # browse and switch TTS voices
rcli cleanup         # remove unused models

Practical Evaluation Checklist

• Installs on macOS 13+ Apple Silicon without cloud credentials
• rcli setup completes and reports model download success
• rcli TUI opens, push-to-talk records and returns a response
• rcli ask "what time is it?" executes in single-shot mode
• rcli rag ingest ./docs/ indexes documents without error
• rcli rag query [question] returns grounded answers
• rcli actions lists available macOS actions (40+)
• M1/M2 Mac: falls back to llama.cpp engine automatically
• rcli info reports MetalRT active on M3+

Security Notes

• All inference runs locally — no audio or text is sent to any cloud service
• No API key required; no telemetry or phone-home documented in the README
• Models are downloaded once and stored locally at ~/Library/RCLI/models
• MetalRT is closed-source; if you require full source transparency, build from source with llama.cpp

FAQ

Q: Is RCLI fully local? A: Yes. Every step — STT, LLM inference, TTS, and VLM — runs on your Mac’s Apple Silicon GPU. Nothing is sent to a server.

Q: Does it work on Intel Macs? A: The README does not document Intel Mac support. The CPU-only build (llama.cpp + sherpa-onnx) should run on Intel, but this is not the primary target.

Q: Can I use my own models? A: Yes. rcli models lets you browse, download, and switch LLMs. VLM models can be managed via rcli models vlm.

Q: Is it free? A: RCLI itself is open-source under MIT. MetalRT (the proprietary GPU engine) is included in the pre-built binary at no additional cost, per the GitHub releases.

Q: What open-source license does RCLI use? A: MIT License, verified via the LICENSE file on GitHub.

Conclusion

RCLI stands out in the local AI tooling space by shipping a polished, all-in-one voice interface instead of requiring users to wire together separate STT, LLM, and TTS components. The proprietary MetalRT engine is a meaningful differentiator for M3+ Mac users who want GPU-accelerated inference without cloud dependencies. The MIT license and active development (last release March 2026) make it worth watching.