What it does OpenJarvis is a Python framework for building personal AI agents that run primarily on your own hardware. It bundles Ollama, a Rust extension, and eight built-in agents — from a morning digest that reads your email and calendar aloud, to a code-execution agent, to continuous monitoring agents with persistent memory. You install with one curl or PowerShell command, then drive it through a jarvis CLI.

The interesting bit The project treats “Intelligence Per Watt” as a first-class metric alongside accuracy. It comes from Stanford’s Scaling Intelligence Lab and explicitly aims to be “PyTorch for local AI” — a research platform that doubles as production infrastructure. The learning loop that improves models from local trace data is still early, but the ambition is unusual: optimize for energy, FLOPs, latency, and dollar cost, not just benchmark scores.

Key highlights

• One-liner installers for macOS, Linux, WSL2, native Windows, plus desktop GUI binaries (.exe, .dmg, .AppImage, etc.)
• Skills system with ~13,700 importable community skills via OpenClaw, following the agentskills.io open standard
• Eight built-in agent types across three execution modes: on-demand, scheduled, and continuous
• jarvis doctor for diagnostics; jarvis bench for benchmarking skill impact
• Apache 2.0, with active roadmap and auto-issue-assignment via “take” comments

Caveats

• The “learning loop” and local trace-based optimization are described aspirationally; current capabilities appear centered on the CLI agent framework and presets
• Some features (Rust extension, larger models) download after initial install, so first-run latency may surprise
• Windows native support requires scheduled-task service setup; WSL2 has its own notes

Verdict Worth a look if you want a structured, hackable local-agent framework with Stanford research credibility. Skip it if you need mature, fully autonomous learning today — or if you’re happy paying OpenAI by the token and don’t mind the privacy trade.