OpenJarvis - Stanford University's open-source native AI agent framework

What is OpenJarvis?

OpenJarvis is an open-source, local AI agent framework developed by Stanford University's Scaling Intelligence Lab. Its core concept is to make AI execution completely localized, with cloud access as an option. The framework provides five main modules: a unified model catalog layer, a hardware-aware inference engine, an agent orchestration system, tool memory, and learning optimization. It supports various methods... pip install openjarvis One-click installation offers four interaction methods: browser, desktop application, Python SDK, and CLI. Research shows that the local model can handle 88.7% of single-turn dialogue queries, with intelligent efficiency improving 5.3 times in two years. OpenJarvis evaluates energy consumption, latency, cost, and accuracy equally, and includes standardized evaluation tools, making it suitable for privacy-sensitive individual users and providing a feasible path for enterprises to reduce their reliance on cloud APIs.

Main functions of OpenJarvis

• Local-first architectureAI execution runs entirely on personal devices, with cloud access being optional, and data remaining locally by default to protect privacy.
• Unified model managementThe Intelligence module provides a local AI model catalog layer, simplifying model selection, compatibility, and deployment processes.
• Hardware-aware inferenceThe Engine module automatically detects hardware configuration and supports multiple inference backends such as Ollama, llama.cpp, vLLM, and SGLang.
• Intelligent agent orchestration systemThe Agents module translates model capabilities into executable behaviors, supporting the division of labor between Orchestrator and Operative roles.
• Tools and memory connectionIt supports semantic indexing and persistent storage of local content through the MCP protocol and Google A2A connection to external tools.
• Local learning optimizationThe Learning module performs closed-loop optimization based on local interaction data, supporting fine-tuning methods such as SFT, GRPO, and DPO, as well as automatic optimization of prompt words.
• Standardized efficiency evaluation:built-in jarvis bench The tool evaluates energy consumption, latency, FLOPs, dollar cost, and accuracy equally.
• Multi-platform interactionIt offers four usage methods: browser application, desktop application (macOS/Windows/Linux), Python SDK, and CLI.
• OpenAI compatible API:jarvis serve Start the service to provide a local API endpoint compatible with the OpenAI format.

OpenJarvis's technical principles

• Local priority execution modelIt adopts a "cloud as an option" architecture, and runs inference on personal devices (laptops/workstations) by default. 88.7% of single-turn dialogues and inference queries do not require cloud intervention, and achieve efficient local computing through hardware accelerators (Apple Silicon, NVIDIA GPUs, etc.).
• Modular layered architectureThe system features a loosely coupled design of five modules: Intelligence (model abstraction layer), Engine (inference runtime), Agents (behavior orchestration), Tools & Memory (external connectivity), and Learning (closed-loop optimization). Each layer can be upgraded or replaced independently.
• Hardware-aware scheduling mechanismThe Engine module automatically detects available hardware resources (CPU/GPU/NPU) and dynamically selects the optimal inference backend (Ollama/llama.cpp/vLLM/SGLang). jarvis init and jarvis doctor Complete environmental diagnosis and configuration recommendations.
• Agent role separationAgents adopts an Orchestrator-Operative division of labor model. The Orchestrator is responsible for task planning and tool selection, while the Operator focuses on executing specific actions. It supports both single-agent and multi-agent collaboration modes.
• MCP/A2A Protocol IntegrationThe Tools & Memory layer connects to external data sources and tools through the Model Context Protocol and the Google A2A standard protocol to achieve semantic indexing (vector database) and persistent memory storage of local content.
• Local Continuous Learning FrameworkThe Learning module performs local model fine-tuning based on user interaction data, supporting three algorithms: SFT (Supervised Fine-tuning), GRPO (Group Relative Policy Optimization), and DPO (Direct Preference Optimization), as well as Automatic Prompt Optimization (APE).
• Multi-dimensional efficiency evaluation systemBreaking away from the traditional single indicator of accuracy, it establishes a four-dimensional evaluation matrix encompassing energy consumption (Watts), latency (ms), computational load (FLOPs), and cost ($), with built-in... jarvis bench Standardized evaluation tools enable reproducible comparisons.

OpenJarvis project address

• Github repositoryhttps://github.com/open-jarvis/OpenJarvis

Application scenarios of OpenJarvis

• Privacy-sensitive personal assistantSuitable for professionals such as lawyers, doctors, and psychologists who handle confidential information. All data is stored locally, avoiding the risk of sensitive content leakage.
• Offline environment smart officeBusiness travelers who frequently travel or are in environments with unstable internet connections can continue to use AI to complete document processing, email drafting, and schedule planning even in offline scenarios such as airplanes and high-speed trains.
• Local knowledge base Q&AEmployees import internal documents, code repositories, and product manuals into a local vector database to build a private intelligent search and question-answering system, replacing traditional knowledge management tools.
• AI development for edge devicesIoT developers are using OpenJarvis to deploy lightweight agents on edge devices such as Raspberry Pi and Jetson, enabling real-time inference scenarios such as smart home control and industrial quality inspection.
• Low-cost AI prototypingStartups and individual developers can significantly reduce token consumption costs during the development phase by using local models instead of cloud API calls, and support a smooth migration from prototype to production.
• Educational research experimental platformUniversity researchers utilize the framework's built-in... jarvis bench Evaluation tools and various fine-tuning methods were used to conduct academic experiments on model efficiency optimization and agent behavior research.