OpenJarvis is a framework for local-first personal AI agents that run on your own hardware and call cloud APIs only when needed. It comes from Stanford research on intelligence per watt, which found that local language models already handle most single-turn chat and reasoning queries, and it aims to be both a research platform and a foundation for building on-device agents.
The framework is built around five primitives. Models: it picks or manages local models across providers. Engines: it runs inference on Ollama, vLLM, SGLang, llama.cpp, or cloud APIs, and auto-detects the hardware to recommend the best fit across Apple Silicon, NVIDIA, AMD, NPUs, and CPUs. Agents: multi-step reasoning with tool use. Tools and memory: web search, a calculator, file I/O, a code interpreter, retrieval, persistent local state, and any MCP server. Learning: interaction traces drive automatic improvements to weights, prompts, and agent behavior.
Evaluations treat energy, FLOPs, latency, and dollar cost as constraints alongside accuracy. OpenJarvis ships eight built-in agents across on-demand, scheduled, and continuous modes, plus starter presets such as a spoken morning digest from email and calendar, deep research with citations, and a code assistant. Skills follow the agentskills.io standard and can be imported from Hermes Agent, OpenClaw, or any Git repository, then optimized from trace history. It runs as a CLI, a browser GUI, or a desktop app, and can expose an OpenAI-compatible API.
Features
- Local-first inference: Ollama, vLLM, SGLang, and llama.cpp engines, with cloud APIs as an option
- Hardware detection: recommends a model and engine for Apple Silicon, NVIDIA, AMD, NPUs, or CPUs
- Built-in agents: eight agents including morning digest, deep research, monitoring, orchestrator, ReAct, and CodeAct
- Execution modes: on-demand, scheduled, and continuous agents with persistent state
- Starter presets: morning digest, deep research, code assistant, scheduled monitor, and simple chat
- Tools and memory: web search, calculator, file I/O, code interpreter, retrieval, persistent local state, and MCP servers
- Skills: agentskills.io-format skills imported from Hermes Agent, OpenClaw, or Git repositories, and optimized from traces
- Learning loop: local traces drive improvements to weights, prompts, and agent behavior
- Efficiency-aware evaluation: benchmarks track energy, FLOPs, latency, and cost alongside accuracy
- OpenAI-compatible API: serves local models through an OpenAI-compatible endpoint
- Interfaces: CLI, browser GUI, and desktop apps for macOS, Linux, and Windows
