DEV.co
RAG Frameworks · RunanywhereAI

RCLI

RCLI is a macOS voice assistant that runs entirely on-device using Apple Silicon, combining speech-to-text, LLM inference, text-to-speech, and document Q&A (RAG) without cloud calls. It includes a proprietary GPU engine (MetalRT) for M3+ chips and falls back to llama.cpp on M1/M2, with 40 built-in macOS automation actions.

Source: GitHub — github.com/RunanywhereAI/RCLI
1.5k
GitHub stars
83
Forks
C++
Primary language
MIT
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
RepositoryRunanywhereAI/RCLI
OwnerRunanywhereAI
Primary languageC++
LicenseMIT — OSI-approved
Stars1.5k
Forks83
Open issues12
Latest releasev0.3.7 (2026-03-15)
Last updated2026-03-16
Sourcehttps://github.com/RunanywhereAI/RCLI

What RCLI is

Built in C++, RCLI implements a concurrent STT + LLM + TTS pipeline using Metal GPU acceleration. It supports multiple LLM/STT/TTS model families (Qwen3, LFM2, Whisper, Kokoro), includes hybrid BM25+vector RAG with ~4ms latency, on-device VLM analysis (camera/screen), and tool-calling for macOS automation via AppleScript and shell commands.

Quickstart

Get the RCLI source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/RunanywhereAI/RCLI.gitcd RCLI# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Privacy-First Voice Assistant on macOS

Organizations handling sensitive data (legal, healthcare, finance) can deploy a fully local voice AI without sending audio or documents to cloud APIs.

Document Intelligence & Knowledge Workers

Knowledge workers ingest proprietary documents (PDFs, Word docs) and query them by voice with RAG, supporting multi-turn conversations without external indexing services.

M3+ Mac Automation & Task Orchestration

Teams needing local macOS automation (Spotify control, note creation, app launching, shortcuts execution) can voice-command 40 built-in actions with sub-200ms latency.

Implementation considerations

  • Requires M3 or later Apple Silicon for MetalRT; M1/M2 auto-falls back to llama.cpp with 2–5x slower inference. Verify target Mac model before rollout.
  • Models (~1GB) download during `rcli setup` one-time. Plan for initial network bandwidth and disk space on air-gapped Macs; offline model management is manual.
  • macOS actions use AppleScript and shell commands; 40 actions are hard-coded. Extending actions requires modifying source code or contributes to the repo.
  • RAG ingests local files only (PDF, DOCX, TXT); no built-in sync with cloud storage (OneDrive, S3, etc.). Document updates require manual re-ingest.
  • No user/session management, audit logging, or multi-user access control. Suitable only for single-user or development teams, not regulated environments.

When to avoid it — and what to weigh

  • Requires Linux or Windows Deployment — RCLI is macOS-only (Apple Silicon required). No Windows, Linux, or Intel Mac support. Non-Mac teams must use alternatives.
  • Need Real-Time Multi-User Concurrency — RCLI is a single-user CLI/TUI. No multi-seat licensing, server mode, or API gateway for distributed teams. Not suitable for call centers or shared systems.
  • Require Vendor Long-Term Stability Guarantees — Project is ~2 weeks old (created 2026-03-04), backed by a private company (RunAnywhere, Inc.). No production SLA, no long-term support roadmap documented, no enterprise contracts.
  • MetalRT Dependency for Production Performance — MetalRT (proprietary) is required for <200ms latency and 550 tok/s throughput on M3+. Its license is non-standard and unavailable for public review; falls back to llama.cpp on M1/M2, reducing performance significantly.

License & commercial use

RCLI core is MIT-licensed (permissive, allows commercial use). MetalRT GPU engine is proprietary (non-standard license, requires review with legal). Vision models (Qwen3 VL, LFM2 VL, SmolVLM) licenses not specified; llama.cpp is Apache 2.0.

RCLI source code (MIT) permits commercial use. However, MetalRT binary is proprietary and its commercial use terms are unclear; licensing inquiries directed to [email protected]. Recommend legal review before shipping MetalRT-dependent features in a commercial product.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityNeeds review
Deployment complexityModerate
DEV.co fitPossible
Assessment confidenceMedium
Security considerations

RCLI runs entirely on-device; no remote exfiltration of audio or documents. However, no formal security audit disclosed. Considerations: voice data is not encrypted at rest, AppleScript/shell execution could be exploited if LLM output is not validated, RAG index is plaintext, and no user authentication. Air-gapped Macs mitigate cloud risk but not local privilege escalation.

Alternatives to consider

Apple Siri + Shortcuts + Local ML

Native, supported, but limited tool-calling, weaker NLU, no RAG, and tightly coupled to Apple ecosystem. No fine-tuned LLM control.

Ollama (llama.cpp wrapper) + Whisper + Piper

Open-source, cross-platform alternative to MetalRT. Slower on M3+ but no proprietary license risk. Requires manual orchestration of STT/TTS/LLM pipelines.

Open WebUI + Local LLM (self-hosted on Intel/Mac)

Browser-based, multi-user, supports RAG and model swapping. Slower inference, more deployment overhead, but vendor-agnostic and fully open-source.

Software development agency

Build on RCLI with DEV.co software developers

Test RCLI's on-device voice pipeline on Apple Silicon M3+ for proof-of-concept. Review MetalRT licensing and production readiness (2-week-old project) with your legal and engineering teams before commercial rollout.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

RCLI FAQ

Can I use RCLI on M1 or M2 Macs?
Yes, but MetalRT is unavailable on M1/M2. RCLI automatically falls back to llama.cpp, reducing LLM throughput to ~100–150 tok/s (vs. 550 tok/s on M3+) and increasing voice latency. Performance is acceptable for light use but not production.
Is MetalRT open-source?
No. MetalRT is proprietary and distributed as a binary. Source code is not available. License is non-standard and not publicly reviewed; contact [email protected] for commercial licensing.
Can I extend the 40 macOS actions?
Not without modifying source code or contributing to the GitHub repo. Actions are hard-coded. Extending them requires C++ development and a pull request.
Does RCLI support multi-user or server deployment?
No. RCLI is a single-user CLI/TUI tool. No multi-seat licensing, server mode, API gateway, or session management. Not suitable for teams or shared infrastructure.

Work with a software development agency

Adopting RCLI is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate rag frameworks software in production.

Evaluate RCLI for Your macOS Privacy Needs

Test RCLI's on-device voice pipeline on Apple Silicon M3+ for proof-of-concept. Review MetalRT licensing and production readiness (2-week-old project) with your legal and engineering teams before commercial rollout.