OpenJarvis 0.1.2.dev753 Enables On‑Device AI with Optional Cloud

OpenJarvis 0.1.2.dev753 is a research‑grade framework that lets developers build personal AI agents that run on their own hardware rather than relying on remote servers [1]. The platform bundles a browser‑based UI, a native desktop app, and a Python SDK, all of which connect to a locally hosted inference backend such as Ollama, vLLM, or llama.cpp [2].

Key takeaways

• OpenJarvis runs entirely on the user’s device; cloud services are optional [1]
• A single script can install dependencies, start a local model, and launch the UI at http://localhost:5173 [2]
• The framework supports five composable layers—Intelligence, Engine, Agents, Tools & Memory, and Learning—allowing interchangeable components [1]
• Multiple inference backends are supported, with Ollama recommended for most consumer hardware [2]
• Built‑in telemetry tracks energy use, token cost, and latency for each query [1]

Local‑First Architecture and Extensible Components

OpenJarvis is built around a “five‑primitive” design that separates model selection, inference runtime, agent logic, tool integration, and continuous learning [1]. Users can pick a model manually or let the system choose one that matches their GPU’s vendor, model, and VRAM, while the Engine layer automatically selects the optimal runtime—Ollama, vLLM, SGLang, llama.cpp, or even cloud APIs if desired [1]. Agents range from simple chat bots to orchestrated workflows that can invoke tools such as web search, calculators, or file I/O, and the Learning layer records interaction traces to improve prompts and model weights over time [1].

Installation is streamlined for both browser and desktop experiences. The “one‑command” setup checks for Python 3.10+ and Node.js 18+, installs Ollama if needed, pulls a starter model, and launches both backend and frontend servers [2]. For users who prefer manual control, the repository provides step‑by‑step commands to clone the code, sync dependencies with uv, and build the Rust components via maturin [2]. The desktop app, available for macOS, Windows, and Linux, opens a native window that connects to the locally running backend at http://localhost:8000 [2].

Tooling, Telemetry, and Community Support

Beyond the core AI stack, OpenJarvis includes a Python SDK that offers a high‑level synchronous API for programmatic queries, complete with support for agents, tools, and telemetry [2]. Example code shows how to ask a question, invoke specific tools like a calculator, and retrieve turn counts and tool results [2]. The platform also provides a CLI for quick interactions, and a FastAPI server (jarvis serve) that mimics OpenAI’s API, enabling drop‑in compatibility for existing client libraries [1].

OpenJarvis is part of the “Intelligence Per Watt” research initiative, a collaboration between Hazy Research, Stanford’s Scaling Intelligence Lab, and sponsors including Cloud Platform and IBM Research [1]. The project’s documentation covers installation guides, Docker images, systemd launch scripts, and a roadmap for future development [1].