Jarvis is a free, MIT-licensed, open-source voice dictation and voice-control assistant for the Mac, built for anyone who wants hold-a-key dictation and local, private AI processing without a subscription.

What it is

Jarvis is a TypeScript desktop application for macOS that turns speech into clean, punctuated text. The workflow is deliberately small: hold the Fn key, speak, release, and formatted text appears in whatever application has focus. It lives in the Productivity & Utilities / Input & Dictation category and covers both dictation and lightweight assistant behaviour, so the same keystroke can clean up a rambling sentence, rephrase it as a bullet list, or trigger a short action such as "open YouTube" or "set 5-min timer".

The concrete problem it solves is the paid dictation subscription. The project's README opens by noting that Wispr Flow raised $81 million to build a voice dictation app, and Jarvis exists as one developer's open-source answer to that. Where the commercial path charges roughly $10–24 per month and offers no offline mode, Jarvis is free forever, runs under the MIT licence, and can transcribe entirely on-device through local Whisper or NVIDIA Parakeet models. Speech and text never have to leave the machine, and the README states plainly that there is zero tracking and zero telemetry.

Key capabilities

• Hold Fn to record, double-tap Fn to toggle hands-free mode, and press Escape to cancel a recording.
• Offline transcription with local Whisper (tiny, base, or small) or NVIDIA's 600M-parameter Parakeet model via Sherpa-ONNX.
• Local LLM post-processing through Ollama, with automatic model detection; the README shows ollama pull sam860/LFM2:1.2b and ollama pull llama3 as examples.
• Cloud transcription and generation using Deepgram plus Gemini, with a free tier of 1M tokens per day; Deepgram requests default to mip_opt_out=true.
• Automatic filler removal for words such as "um" and "like", plus grammar correction, rephrasing, bullet-point formatting, and text generation.
• Full prompt engineering: every behaviour, including mail formatting, dictation cleanup, and assistant responses, is customizable through prompts and custom voice commands.
• Distribution as signed and notarized DMGs for Apple Silicon (M1/M2/M3/M4) and Intel x64, plus an iOS build on TestFlight.

Who uses it and how

• Mac users on Apple Silicon who download the M1/M2/M3/M4 DMG and run dictation across any application without granting a cloud service their audio.
• Privacy-conscious users who pick the "local-only" onboarding path, enabling a local model under Settings → Transcription so no API keys are required at all.
• Intel Mac owners who install the x64 DMG and get the same dictation workflow on older hardware.
• Users who prioritise speed or accuracy over strict locality, pasting free Deepgram and Gemini keys and accepting a cloud round trip.
• Developers who build from source with Node.js 18+, macOS 10.13+, and Xcode CLI tools, running npm install && npm run build && npm run dev.

Getting started

Download the signed and notarized DMG from the GitHub releases page, choosing the Apple Silicon or Intel x64 build, then enable a local model under Settings → Transcription or paste Deepgram and Gemini keys. Local LLM support additionally requires installing Ollama separately and enabling "Use Ollama" under Settings → AI Models.

How it compares

The README sets Jarvis directly against a paid dictation product priced at roughly $10–24 per month and against the broader "$700M startups" framing, contrasting them on licence (MIT versus proprietary), offline capability (local Whisper and Parakeet versus none), telemetry (none versus undisclosed), and cost model (free forever versus subscription). The most concrete difference is data ownership: with a local model enabled, audio and text stay on the machine rather than being sent to a vendor.

When to use it — and when not to

Choose Jarvis if the machine is a Mac running macOS 10.13 or later and the priority is free, auditable, offline-capable dictation. Be aware of the operating costs: local LLM post-processing adds roughly one to three seconds of latency on typical M1 and M2 chips, and faster results on standard hardware mean using Deepgram with Gemini or local Whisper with Gemini Flash instead. Windows users should look elsewhere, since a Windows version remains an unchecked roadmap item, and anyone unwilling to install Ollama or manage API keys should expect a narrower local setup.

git clone https://github.com/akshayaggarwal99/jarvis-ai-assistant.git
cd jarvis-ai-assistant
npm install && npm run build && npm run dev