Stanford drops OpenJarvis for local-first personal AI

OpenJarvis is an open-source framework from Stanford researchers that lets you build and run personal AI agents entirely on your own hardware, using the cloud only as a last resort.

Most modern personal AI projects—like OpenClaw and its ecosystem of spin-offs—are essentially thin local wrappers around cloud APIs. Your most intimate personal data, your context, and your wallet are at the mercy of remote data centers. OpenJarvis wants to change that architecture completely.

What it does

The Stanford team ran an extensive benchmark spanning over 1 million queries, 20 different models, and 8 hardware accelerators. Their finding: local language models and on-device accelerators can accurately handle 88.7% of single-turn chat and reasoning queries at interactive latencies.

That means nearly nine out of ten tasks never need to touch the cloud. On top of that, their Intelligence Per Watt study found that intelligence efficiency jumped 5.3x between 2023 and 2025, proving local chips are finally fast and power-efficient enough for daily workloads.

How it works

OpenJarvis breaks on-device AI into five modular primitives: Intelligence, Engine, Agents, Tools & Memory, and Learning. You can plug into the whole stack or benchmark each component individually.

Two things make it stand out. First, it treats efficiency as a first-class citizen. The built-in evaluation tools measure energy, latency, FLOPs, memory footprint, and dollar cost alongside raw accuracy, which is essential when running on laptop batteries.

Second, it introduces a closed-loop optimization harness. The system can fine-tune prompts, agentic logic, and weights using local session trace data, letting your assistant improve over time without sending a single token back to an external server.

The framework is available now for download, and the team is even running an energy efficiency leaderboard contest where top contributors can win a Mac Mini.

Why it matters

If you care about local privacy, hate getting hit with token bills, or want to build an agent that works completely offline, OpenJarvis is well worth setting up.

We think its real value is showing the math behind local AI: 88.7% of what we ask personal models to do simply doesn't require a remote cluster. By turning scattered local scripts into a standardized, PyTorch-like stack, Stanford is laying a solid foundation for agents that respect both your battery and your privacy.