OpenJarvis Brings Fully Local‑First Personal AI to Your Device

Overview

The OpenJarvis project, hosted on GitHub, introduces a local‑first software stack for personal AI agents. While most consumer‑grade assistants still rely on cloud APIs, OpenJarvis aims to run the bulk of inference on‑device, reaching out to the cloud only when a task truly requires external resources. The initiative is part of the *Intelligence‑Per‑Watt* research program run by Hazy Research and Stanford’s Scaling Intelligence Lab.

Core Ideas

OpenJarvis is built around three principles that the repository documentation highlights:

• Shared primitives – a common set of building blocks that let developers create on‑device agents without rewriting low‑level code.
• Energy‑aware evaluation – performance metrics such as FLOPs, latency, power draw and cost are treated as first‑class constraints alongside traditional accuracy measures.
• Local learning loop – agents can improve over time by ingesting trace data generated on the user’s hardware, reducing the need for frequent cloud retraining.

The goal, as stated by the project, is to make “personal AI agents that run locally by default, calling the cloud only when truly necessary.”

Installation & Quick Start

The repository provides platform‑specific installers that automate most of the setup:

1. Choose your OS (Linux, macOS, Windows WSL2, etc.).
2. Run a single command; the script creates a Python virtual environment, installs uv, pulls Ollama, and downloads a starter language model (≈3 minutes on a broadband connection).
3. Launch the assistant with `jarvis`.
4. A Rust extension and larger models continue downloading in the background; `jarvis doctor` reports status.

For a graphical interface, `jarvis gui` starts a local API server, launches a Node.js‑based front‑end, and opens the chat window in the default browser. The GUI requires Node 22+ and is available from source or as a packaged desktop installer.

Built‑In Agents & Skills

OpenJarvis ships with eight pre‑configured agents that operate in three modes:

• On‑demand – invoked manually.
• Scheduled – run at defined intervals.
• Continuous – always listening for triggers.

Agents acquire *skills*—discrete tool definitions—from a public catalog (e.g., Hermes Agent, OpenClaw) or any GitHub repository that follows the `agentskills.io` standard. Skills enable agents to perform tasks such as:

• Morning news digests
• Deep‑research summarisation
• Code assistance
• System monitoring

Community, Roadmap & Contributions

The project encourages external contributions. A Contributing Guide outlines incentive structures, PR processes, and how to auto‑assign issues by commenting “take.” The public Roadmap lists priority areas where help is needed, ranging from Docker deployment to new skill authoring.

Verifying the Claims

All technical details come directly from the OpenJarvis GitHub repository (https://github.com/open-jarvis/openjarvis) and its accompanying documentation site (https://open-jarvis.github.io/OpenJarvis) Interested readers can:

• Review the `README.md` and `docs/` folder for architecture diagrams.
• Clone the repo and run the installer to test the on‑device inference claim.
• Examine the `benchmarks/` directory for the cited *Intelligence‑Per‑Watt* efficiency numbers (88.7 % single‑turn coverage, 5.3× improvement projected to 2025).

Practical Implications

If the stack lives up to its promises, developers and power users could:

• Deploy privacy‑preserving assistants that never send raw user data to external servers.
• Reduce operational costs by limiting cloud compute usage.
• Experiment with custom agents tailored to niche workflows without needing a large‑scale backend.

--- *The article reflects information publicly available on the OpenJarvis GitHub page as of November 2026. No independent testing has been performed.*