Stanford unveils OpenJarvis, a stack for local AI agents with memory and learning

Stanford researchers have released OpenJarvis, an open-source framework for personal AI agents that run entirely on the user's device. The project is designed as a ready-made stack for local AI, covering everything from model execution and agent orchestration to memory, tools, benchmarks, and follow-up training on local data.

Why it matters

Most personal AI systems are still local only on the surface: the interface runs on a laptop, but the heavy reasoning goes out to cloud APIs. For tasks involving files, notes, email, and persistent user context, that means latency, recurring costs, and unnecessary transfers of sensitive data. OpenJarvis offers a different model: local execution by default, with the cloud as an option only when it's genuinely needed.

Stanford ties the release to its own Intelligence Per Watt research. According to the lab, local language models and local accelerators can already correctly handle 88.7% of single-turn chat and reasoning queries at interactive response speeds, and efficiency measured as "intelligence per watt" improved 5.3x between 2023 and 2025. The premise behind OpenJarvis is that the hardware and models are nearly ready, and what the market has lacked is a unified software layer for these systems.

How the stack works

OpenJarvis is built around five primitives that can be swapped, tested, and optimized independently. The approach aims to eliminate the mess typical of local AI setups, where inference, agent logic, tool use, memory, and training are tangled into a single project that's hard to reproduce. Developers can benchmark a specific layer (the model, engine, memory, or agent behavior) instead of the whole system at once. That makes both experimentation and production deployment noticeably easier.

• Intelligence — the model layer, with a unified catalog of local LLMs and an abstraction for choosing between them.
• Engine — the runtime, which runs models via Ollama, vLLM, SGLang, llama.cpp, and other engines.
• Agents — agent roles, including an Orchestrator that breaks down tasks and an Operative for recurring workflows.
• Tools & Memory — access to tools, local memory, semantic search, MCP, and agent-to-agent communication via A2A.
• Learning — an improvement loop that uses local traces for fine-tuning and optimization.

The team put particular emphasis on making the system more than a chatbot. OpenJarvis can run local search across notes and documents, hook into tools such as web search, a calculator, file I/O, and code interpretation, and connect to external MCP servers. As a result, the framework is pitched not as a wrapper around a single model but as infrastructure for a personal agent with long-term memory and access to the user's real environment.

What's available now

On the practical side, the project is fairly down-to-earth. OpenJarvis offers a CLI, a Python SDK, a browser interface, and a desktop app for macOS, Windows, and Linux. The `jarvis init` command detects available hardware and recommends a suitable engine-and-model pairing, `jarvis doctor` helps diagnose the configuration, and `jarvis serve` spins up an OpenAI-compatible API server on FastAPI, so developers can plug in existing clients and frontends with almost no rework. According to the documentation, basic workflows can run with no network connection at all.

Another strength is that it measures efficiency, not just answer quality. The framework collects telemetry on energy, latency, FLOPs, and the dollar cost per query, supports profiling on NVIDIA, AMD, and Apple Silicon, and standardizes benchmarks through `jarvis bench`. OpenJarvis also stores local interaction traces, from prompt-completion pairs to sequences of agent actions and tool calls. Developers can use that data to optimize not only model weights but also prompts, agent logic, and the inference engine itself, for example via quantization, DSPy, GEPA, SFT, DPO, or GRPO.

What it means

OpenJarvis shows local AI moving out of the experimental-build category and toward a full engineering stack. If Stanford's approach catches on, developers will get a standard foundation for personal agents that keep data with the user, cost less to run, and become more useful over time by learning from the user's own local workflows. For the market, it's another signal that some everyday AI tasks will soon start migrating from the cloud to personal devices.