Stanford's OpenJarvis Runs Personal AI Agents 800x Cheaper Without the Cloud
Stanford's OpenJarvis is a free, open-source framework that runs full AI agents locally on your hardware, with no cloud required and a built-in learning loop
Takeaways
- What it is: OpenJarvis is a free, open-source framework from Stanford for running full personal AI agents locally on your hardware.
- Key claim: Local models already handle 88.7% of single-turn queries; OpenJarvis provides the missing software stack to make this practical.
- Architecture: Five composable primitives -- Intelligence, Engine, Agents, Tools & Memory, and Learning -- each independently swappable via a TOML config.
- Performance: Within 3.2pp of best cloud models on benchmarks, at ~800x lower API cost and ~4x lower latency.
- Ships with: 8 built-in agents, 13,700+ community skills via OpenClaw, MCP tool support, and a closed-loop local learning loop.
- Caveats: The desktop frontend has routing/onboarding issues; memory writes require explicit API calls, not casual chat instructions.
OpenJarvis is a new open-source framework from Stanford's Hazy Research and Scaling Intelligence Lab that lets you run personal AI agents entirely on your own hardware. It is free, Apache 2.0 licensed, and already has 6,800+ stars on GitHub. The pitch is simple but pointed: most "personal" AI today is just a thin wrapper around someone else's cloud API. OpenJarvis is the software stack to actually change that.
The problem it's solving
Personal AI agents are exploding in popularity, but nearly all of them still route intelligence through cloud APIs -- your "personal" AI continues to depend on someone else's server. The Stanford team's own prior research gave them a concrete reason to believe this is fixable: their Intelligence Per Watt study found that local models already handle 88.7% of single-turn queries. The models and hardware are ready. What was missing was the software stack to make it practical.
OpenJarvis is built around three core ideas: shared primitives for building on-device agents; evaluations that treat energy, FLOPs, latency, and dollar cost as first-class constraints alongside accuracy; and a learning loop that improves models using local trace data. The goal is to make it possible to build personal AI agents that run locally by default, calling the cloud only when truly necessary.
Five composable building blocks
The architecture is organized into five independently swappable primitives, composed through a single declarative TOML config. This is the key design choice: each layer can be benchmarked and optimized on its own, or studied as part of the full system.
- Intelligence -- the model layer. It sits above a changing set of local model families and provides a unified model catalog so developers do not have to manually track parameter counts, hardware fit, or memory tradeoffs for every release. Supported families include Qwen, Gemma, Granite, GLM, and others.
- Engine -- the inference runtime. It is a common interface over backends such as Ollama, vLLM, SGLang, llama.cpp, and cloud APIs. Commands like