OpenJarvis: local AI that isn't quite local yet

Claim

Personal AI, on your device?

Catch

RESEARCH, NOT A FINISHED PRODUCT

Verdict

Research stack, not appliance

OpenJarvis pitches itself as "personal AI on personal devices" — a framework from Stanford's Hazy Research and Scaling Intelligence Lab for agents that run locally by default and call the cloud only when necessary. It has 12,175 stars on GitHub, the Apache 2.0 license, and a one-command installer. On paper it's exactly what local-AI enthusiasts want.

The claim

The pitch is that you can build personal AI agents on your own hardware, keeping your data and your model on device. The README lists eight built-in agents (morning digest, deep research, code assistant, scheduled monitor, and more) and a jarvis CLI plus a jarvis gui that opens a browser chat. The installer handles uv, Ollama, and a starter model in about three minutes on broadband.

What the paper actually says

The claim that local models can stand in for cloud frontier models doesn't hold up out of the box. The OpenJarvis paper (arXiv:2605.17172, submitted 16 May 2026) shows that swapping a cloud frontier model for a local one drops accuracy 25–39 percentage points on personal-AI tasks like PinchBench and GAIA, and prompt optimization alone closes just 5 points. So "local-first" is a real engineering problem, not a free swap.

OpenJarvis's answer is LLM-guided spec search: frontier cloud models propose edits across a five-primitive spec (Intelligence, Engine, Agents, Tools & Memory, Learning) at search time, only non-regressing edits are accepted, and the resulting spec runs entirely on-device at inference. The result: on-device specs match or beat cloud accuracy on 4 of 8 benchmarks and land within 3.2 points of the best cloud baseline on average, cutting marginal API cost roughly 800×.

The catch

Here's where the marketing runs ahead of the product:

• It's a preprint, not a finished product. The paper is dated May 2026. The README itself describes early-stage code. Treat it as a research platform, not a plug-and-play appliance.
• Bigger models download in the background. The installer grabs uv, Ollama, and a starter model, but larger models keep downloading after you start; jarvis doctor shows status.
• The GUI needs Node.js 22+. jarvis gui starts a local API server and frontend, but it requires Node 22 and is available from a source checkout; packaged .exe/.dmg/.deb/.rpm/.AppImage installers come from the GitHub releases page.
• Local-first can still call the cloud. The whole point of spec search is to decide when the cloud is genuinely needed — so your data may leave your device even though the default is local.

How to try it

Check the package name against the README before installing — typosquats on popular names are common. The official installer is install.sh from open-jarvis.github.io/OpenJarvis, not a random npm or pip package.

# macOS / Linux / WSL2
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash

# then
jarvis
jarvis gui        # needs Node.js 22+

Undo: OpenJarvis installs into a Python venv and Ollama on your machine; remove the venv folder and the Ollama models directory to uninstall. Nothing touches your system outside the scratch-like install location unless you configure otherwise.

The short version: OpenJarvis is a promising research stack for local-first personal AI, and its numbers are real — but it's research code from May 2026, not a finished product you can hand to a non-technical user today.

Sources: github.com/open-jarvis/OpenJarvis; arXiv:2605.17172 (DOI 10.48550/arXiv.2605.17172); open-jarvis.github.io/OpenJarvis