isair/jarvis
Productivity & UtilitiesA local-first desktop AI assistant with always-on voice, personal memory, MCP tools and dictation for macOS, Windows and Linux.
Overview
Jarvis is a desktop AI assistant that listens for its name during nearby conversation, keeps a short rolling transcript, and can weigh in on what was just discussed. Chat, speech recognition and speech synthesis run on hardware you control via a local model server, and conversations, diary entries and the knowledge graph stay on the machine.
The assistant ships with built-in tools for web search, weather, time, screenshot OCR, file access, optional text/Markdown folder search, nutrition tracking and optional location awareness. Additional capabilities come from MCP servers, so browser automation, smart-home control and similar integrations can be plugged in. A companion chat window shares the same conversation and memory for text replies, and a dictation hotkey pastes transcribed speech into other applications. An optional Background Tasks setting, off by default, queues work locally.
Quick start
- Download the package for your platform from GitHub Releases:
Jarvis-macOS-arm64.zip/Jarvis-macOS-x64.zip,Jarvis-Windows-x64.zip, orJarvis-Linux-x64.tar.gz. - Unpack it. On macOS, move the app to Applications and right-click → Open; on Windows, run
Jarvis.exe; on Linux, run./Jarvis/Jarvis. - Follow the setup wizard to pick a speech recognition model and a model server — Ollama or any OpenAI-compatible local server such as LM Studio, oMLX or llama.cpp.
- Grant microphone access and wait for the first model downloads to finish. Once it reports that it is listening, say "Jarvis, help me think through my day".
Right-click the floating face or use the tray menu to open Chat and get silent text replies. On Apple Silicon, unified memory is shared with the system; other GPUs use dedicated VRAM, and setup estimates the budget for chat, fast routing and Whisper models together. Smaller models such as qwen3.5:0.8b trade capability for lower resource use.
Constraints and limitations
- Development is primarily on macOS; Windows and Linux behaviour may differ, and tested capabilities are documented in the evaluation results.
- Accuracy and speed depend on the chosen models and hardware; slow speech recognition can drop utterances.
- Dictation is unavailable on macOS 26+ (#172); on Linux it requires X11 with limited Wayland support.
- A spoken "stop" can be mistaken for echo while Jarvis is speaking (#24), and there is no mobile app (#17).
- Most settings live in the tray Settings UI; advanced options are edited in
~/.config/jarvis/config.json. The configuration guide covers model servers, speech, MCP integrations and troubleshooting. - Memory is stored under
~/.local/share/jarvis. Sensitive information is redacted before reaching model context and the saved diary, but the in-memory chat still shows what was typed. Model downloads, web tools, MCP servers and an external model endpoint can use the network; disabling web tools and settingmcpsto empty is not a network firewall. - The app is free for personal use; commercial use requires contacting the author.
Run from source
Source installs require Python 3.11 or later:
git clone https://github.com/isair/jarvis.git
cd jarvis
# macOS
bash scripts/run_macos.sh
# Windows (PowerShell, with Micromamba)
pwsh -ExecutionPolicy Bypass -File scripts\run_windows.ps1
# Linux
bash scripts/run_linux.sh
Running from source also enables Chatterbox TTS; Piper works in both packaged and source installations.
HotRepo's take
On the Python daily list for 2026-11-16, Jarvis recorded 26 new stars, reaching 7th place with 2,305 stars and 549 forks overall. There is no record of it on the all-language daily list, and this single snapshot does not establish a longer trend. The project's distinguishing position is the combination of an always-listening conversation mode, local model execution and optional dictation in one desktop app. The honest limits — model-dependent reliability, macOS-first development, open dictation issues on macOS 26+ and Wayland — matter more for evaluating fit than the star count. The repository also distributes prebuilt desktop binaries, so source builds or the published releases are the paths for users who want to inspect what they run.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- adewaskar/jarvisA browser-based voice assistant with an Iron Man interface that uses Claude Code to run your tools and MCP servers.
- CapriChooms/CapriCoreA Windows desktop AI companion with local or API models, voice, memory and permission-based desktop tools.
- jouaaaaaa/mcd-self-butlerA McDonald's ordering assistant for people juggling price, nutrition and time, built on the chain's MCP server.
- FrigadeHQ/yogurtRecord meetings locally, transcribe them and generate private AI meeting notes without adding a bot.
- anomalyco/hexLocal speech-to-text dictation app for macOS, Windows and Linux that pastes transcribed speech into any app.