OpenJarvis is a framework for local-first personal AI agents that run on the user's own device by default and call the cloud only when needed. It is developed at Hazy Research and the Scaling Intelligence Lab at Stanford as part of the Intelligence Per Watt research initiative, and it is intended to serve as both a research platform and a production foundation for local AI.
The framework is built around three ideas: shared primitives for building on-device agents; evaluations that treat energy, FLOPs, latency, and dollar cost as first-class constraints alongside accuracy; and a learning loop that improves models from local trace data. The motivating research finding is that local language models already handle 88.7% of single-turn chat and reasoning queries, with intelligence efficiency improving 5.3 times between 2023 and 2025.
It runs on macOS, Linux, WSL2, and native Windows, from the command line or as a desktop GUI, with Ollama and a starter model set up locally and a Rust extension and larger models fetched in the background. The jarvis command starts a chat, jarvis init --preset switches between starter configurations, and jarvis connect gdrive links Gmail, Calendar, and Tasks through one OAuth flow.
Skills teach agents how to use tools and reason better, and every skill is itself a tool that agents discover from a catalog and invoke on demand. Skills follow the agentskills.io open standard and can be imported from Hermes (around 150 skills), OpenClaw (around 13,700 community skills), or any GitHub repository. The CLI can also optimise skills from trace history using a DSPy policy and benchmark the effect.
Features
- Starter presets: morning-digest (spoken briefing from email, calendar, health, and news), deep-research, code-assistant, scheduled-monitor, and chat-simple
- Built-in agents: eight agents across on-demand, scheduled, and continuous modes, including orchestrator, nativereact (ReAct loop), nativeopenhands (CodeAct Python execution), operative, and monitor_operative
- Local inference: Ollama serves models locally by default, with cloud engines available as an option
- Skills catalog: import and sync skills from Hermes, OpenClaw, or GitHub repositories; optimise them from local traces; benchmark the impact
- Efficiency-aware evaluation: energy, FLOPs, latency, and cost measured alongside accuracy, with a public leaderboard
- Learning loop: models improved from local trace data
- Personal data connectors: Google Drive, Gmail, Calendar, and Tasks via a single OAuth flow, plus health and news sources for the digest
- Diagnostics: jarvis doctor reports background download and environment status
