# OpenJarvis

> OpenJarvis is an open-source, local-first personal AI agent framework that runs on-device by default, calling the cloud only when necessary.

OpenJarvis is a local-first personal AI agent framework developed at Hazy Research and the Scaling Intelligence Lab at Stanford SAIL. Released under the Apache 2.0 license, it provides a complete software stack for building and running on-device AI agents across macOS, Linux, Windows, and desktop GUI environments. The project is part of the Intelligence Per Watt research initiative, which studies the intelligence efficiency of AI systems.

## What It Is

OpenJarvis is an open-source framework for personal AI agents that prioritize running locally on the user's own hardware rather than routing through cloud APIs. It is built around three core 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 using local trace data. The project's stated goal is to serve as both a research platform and a production foundation for local AI, in the spirit of PyTorch.

## Architecture and Core Design

OpenJarvis ships with eight built-in agents spanning three execution modes:

- **On-demand agents**: `deep_research` (multi-hop research with citations), `orchestrator` (multi-turn reasoning with automatic tool selection), `native_react` (ReAct loop), `native_openhands` (CodeAct — generates and executes Python code), and `simple` (single-turn chat)
- **Scheduled agents**: `morning_digest` (daily briefing from email, calendar, health, and news with TTS audio)
- **Continuous agents**: `monitor_operative` (long-horizon monitoring with memory, compression, and retrieval) and `operative` (persistent autonomous agent with state management)

The framework integrates with Ollama for local model serving and uses `uv` for Python environment management. A Rust extension handles performance-critical components.

## Skills System

OpenJarvis introduces a skills layer where every skill is a tool that agents discover from a catalog and invoke on demand. Skills can be installed from public sources including Hermes Agent (the README cites approximately 150 skills) and OpenClaw (the README cites approximately 13,700 community skills), or from any GitHub repository. Skills follow the agentskills.io open standard. The framework supports optimizing skills from local trace history using DSPy and benchmarking their impact.

## Setup Path

Installation is designed to be a single command per platform:

- **macOS, Linux, WSL2**: `curl -fsSL ... | bash`
- **Native Windows**: PowerShell one-liner
- **Desktop GUI**: `.exe`, `.dmg`, `.deb`, `.rpm`, or `.AppImage` from the releases page

Each installer handles `uv`, the Python virtual environment, Ollama, and a starter model. The README estimates approximately 3 minutes on broadband. After installation, users run `jarvis` to start and can switch between presets such as `morning-digest-mac`, `deep-research`, `code-assistant`, `scheduled-monitor`, and `chat-simple`.

## Research Backing and Why It Matters

The project is developed at Stanford's Hazy Research group and Scaling Intelligence Lab, with sponsorship from the Laude Institute, Stanford Marlowe, Google Cloud Platform, Lambda Labs, Ollama, IBM Research, and Stanford HAI (as listed in the README). The Intelligence Per Watt research cited in the README found that local language models handle 88.7% of single-turn chat and reasoning queries, with intelligence efficiency improving 5.3× from 2023 to 2025 — a finding the project uses to motivate the local-first approach. The academic paper is available on arXiv (2605.17172).

## Update: Desktop v1.0.2

The latest release as of the GitHub repository data is `desktop-v1.0.2`, published on May 25, 2026. The repository was created in February 2026 and has seen active development, with the last push recorded on June 5, 2026. The project has accumulated over 6,100 stars and 1,300 forks on GitHub according to repository metadata.

## Features
- Local-first AI agent execution
- Eight built-in agents (morning digest, deep research, code assistant, monitor, orchestrator, ReAct, CodeAct, simple chat)
- Skills system with catalog discovery and on-demand invocation
- Skill import from Hermes Agent and OpenClaw community
- DSPy-based skill optimization from local trace history
- Scheduled, on-demand, and continuous agent execution modes
- TTS audio for morning digest briefing
- Multi-hop research with citations
- CodeAct agent for Python code generation and execution
- Long-horizon monitoring with memory, compression, and retrieval
- Single-command installation on macOS, Linux, Windows, and WSL2
- Desktop GUI installers (.exe, .dmg, .deb, .rpm, .AppImage)
- Ollama integration for local model serving
- Google Drive/Gmail/Calendar OAuth connector
- Docker deployment support
- Cloud engine fallback support
- Leaderboard and benchmarking tools
- Energy, FLOPs, latency, and cost as evaluation constraints

## Integrations
Ollama, Google Drive, Gmail, Google Calendar, Google Tasks, Hermes Agent, OpenClaw, DSPy, Docker, uv, agentskills.io

## Platforms
WINDOWS, MACOS, LINUX, CLI

## Pricing
Open Source

## Version
desktop-v1.0.2

## Links
- Website: https://open-jarvis.github.io/OpenJarvis/
- Documentation: https://open-jarvis.github.io/OpenJarvis/
- Repository: https://github.com/open-jarvis/OpenJarvis
- EveryDev.ai: https://www.everydev.ai/tools/openjarvis