The growth of artificial intelligence is moving away quickly off the cloud to personal gadgets. Laptops and Mac now are powerful enough to execute advanced models locally, and provide more personalized, faster, and private AI experiences without turning to remote servers.
RunAnywhere (YC W26) has introduced RCLI, an all-local voice AI platform on macOS, based on their high-performance MetalRT engine. RCLI can comprehend speech, think with documents and run commands on a users computer in near real time all without cloud services, API keys or external infrastructure. Voice-to-action responses take about 131 milliseconds to respond, and so far it is the quickest local voice AI on macOS.
RCLI works solely on-device enabling the user to communicate with their machines in real time, converting spoken instructions directly into actions. This is a significant difference compared to the traditional voice assistants, which in most cases are based on cloud computing, creating latency and privacy risks.
A Fully Local Voice Artificial Intelligence Pipeline.
Voice assistants normally involve the use of several AI elements in collaboration. Speech recognition translates audio into text, a language model understands the query and creates a response and text-to-speech systems create natural spoken output. These elements are used in various frameworks, or rely on cloud APIs in most current solutions, slowing down response time and creating privacy issues.
RCLI operates the whole pipeline locally with MetalRT. Accelerated speech-to-text models are used to first transcribe spoken input into text. The query is then processed by the language model to produce responses or actions in the system. In the case of document-based queries, RCLI uses a hybrid retrieval system that involves HNSW vectors search, BM25 keyword ranking and reciprocal rank fusion (RRF). This enables the assistant to give out applicable results within less than 4 milliseconds with sub-200 milliseconds when embedding caches are used. Lastly, replies are translated into speech using almost real time text-to-speech, and the process is complete in a little more than 1/10 of a second.
This local processing guarantees privacy, speed and reliability, making RCLI a suitable instrument to productivity and entertainment.
Performance Beyond Existing Frameworks.
The performance of RCLI is based on MetalRT, which is optimized to work with Apple Silicon GPUs. The language model inference has 658 tokens per second, nearly 1.19 times faster than MLX on the same hardware. Speed: the Whisper-based models provide staggering speed: audio is transcribed approximately 714 times faster than real-time, 4.6 times faster than MLX Whisper. Text to speech synthesis is also faster with a real time factor of 8.8x, approximately 2.8x faster than MLX Audio and a total voice to audio lag of 63 milliseconds.
In real-life, this would translate to an hour long meeting being transcribed in seconds and answers being instant. Users are presented with a smooth interactive interface that competes with cloud-based assistants yet has data that is stored fully on-device.
Voice Control on MacOS.
In addition to question answering, RCLI offers direct control over macOS by offering an expanding number of 35+ built-in commands. Users are able to start applications, web searches, play media, manage files and interact with services such as Spotify all without internet access.
Due to the open source nature of the system, developers are able to add new commands, making RCLI a fully customizable voice interface to macOS. This flexibility also enables power users to automate workflow, and to control productivity or even run creative applications completely by voice.
Adding Personality to Voice AI
RCLI's latest release introduces Personalities, enabling users to customize how the assistant communicates. The AI can adopt professional, analytical, sarcastic, cynical or even humorously nerdy tones.
This feature allows responses to be tailored to a user's workflow or preference. A professional personality might provide concise, task-focused answers, while a humorous or pop-culture-infused personality can make everyday interactions entertaining. Because the models run locally, experimentation with tone and style is possible without centralized cloud limitations.
Privacy and Offline Operation
Privacy is a key feature of RCLI. All data, including voice recordings, documents and system commands, remains on the user's device. The assistant functions entirely offline, making it ideal for environments with limited internet connectivity or where security is paramount.
Unlike cloud-based assistants that introduce network latency and rely on external servers, RCLI ensures that sensitive information never leaves the user's Mac. Users gain the speed and convenience of a modern voice assistant while retaining full control over their data.
The Future of On-Device Voice AI
RCLI demonstrates the rapid progress of on-device AI. Just a few years ago, running complex AI pipelines locally required specialized hardware or cloud infrastructure. Now, Macs with Apple Silicon processors can perform speech recognition, language reasoning and speech synthesis in real time.
RunAnywhere's MetalRT engine provides the foundation for this shift, optimizing inference for Apple GPUs. With near-instant responses, high-speed transcription and accelerated text-to-speech, RCLI proves that local voice AI can rival cloud systems in speed, accuracy and functionality.
As open-source development continues, the next generation of intelligent assistants may increasingly live entirely on personal devices. RCLI represents an early glimpse of that future: a fully local, open-source voice AI system transforming a Mac into a real-time conversational interface without relying on the cloud.