openjarvis

OpenJarvis is a Python 3.10+ framework and research platform for personal AI agents that run locally by default. Its design emphasizes local models and hardware, with cloud APIs used only when needed. It also treats energy use, FLOPs, latency, and monetary cost as evaluation constraints alongside accuracy, and provides a loop for improving models from local trace data. Installers are provided for macOS, Linux, WSL2, and native Windows, configuring uv, a virtual environment, Ollama, and a starter model. Desktop .exe, .dmg, .deb, .rpm, and .AppImage releases are also offered. The jarvis command starts the system, while jarvis doctor reports background extension and model status. Five starter presets cover lightweight conversation, spoken daily briefings from connected services, multi-hop research with citations, a code agent with execution and file or shell access, and scheduled stateful monitoring. Eight built-in agents operate on demand, on a schedule, or continuously, including simple chat, ReAct reasoning, multi-turn orchestration, code execution, persistent operation, long-horizon monitoring, deep research, and morning briefings with text-to-speech. Skills follow the agentskills.io standard and can be installed from public catalogs or GitHub repositories. Agents discover and invoke them on demand; trace history can be used to optimize skills, and their impact can be benchmarked. The morning-digest preset uses OAuth for services such as Gmail, Calendar, and Tasks and can combine email, calendar, health, and news data. Documentation includes installation, quick-start, preset guides, skills tutorials, Docker deployment, cloud engines, development setup, roadmap, and community links. The project is Apache 2.0 licensed and developed under Stanford's Hazy Research and Scaling Intelligence Lab initiatives.