Stanford just made local AI agents work, and made the cloud look optional

A new open-source framework from Stanford University challenges the idea that personal AI needs the cloud. OpenJarvis, built by the Hazy Research and Scaling Intelligence labs, runs AI agents entirely on the user's own hardware. Cloud connectivity is optional, not mandatory.

Version 1.0, out today, ships with built-in support for Ollama, the local model runner that recently raised $88 million to push open-weight models. The framework sits inside the labs' "Intelligence Per Watt" research agenda, which focuses on making local AI efficient enough for real tasks without hampering performance or draining battery life. The timing is notable: Ollama now reportedly reaches 85% of Fortune 500 companies, according to Ollama's own claims, but whether that number means real adoption or just trial downloads remains an open question.

Local-first by design

Most personal AI assistants, Apple Intelligence, Microsoft Copilot, send every query to remote servers, even for simple tasks like drafting an email or summarizing a meeting. OpenJarvis flips that by routing all inference through the user's own machine by default. The cloud fades to an optional extra instead of the backbone. This distinction matters for users who work offline, for domains where data compliance blocks third-party servers, and for the growing number of developers who want to test models without paying per token. The approach aligns with a broader push for agent infrastructure that lives on the device rather than in a distant data center, as seen with filesystem-based agent frameworks like Vercel's Eve 0.22.

The framework also ships with built-in monitoring that tracks energy consumption, cost, and latency alongside accuracy. Users can see exactly what each query costs in watts and cents, not just in accuracy points. That transparency is a deliberate design choice, one that mirrors recent research showing that simple efficiency metrics often hide real-world performance degradation, a point recently illustrated by DeepSeek's DSpark paper on inference scheduling.

Getting started with Ollama

OpenJarvis ties directly into Ollama, which now runs on macOS, Windows, and Linux. On macOS and Linux, a single curl command detects an existing Ollama installation and sets up the framework. Windows users can run the installer inside WSL2 or download a dedicated desktop app. Once installed, the command-line tool lets users pull any Ollama-compatible model and set a default in a configuration file, making the CLI feel more like a personal assistant and less like an experiment. For newcomers to local models, the learning curve is gentler than expected, especially compared to the plumbing-heavy orchestration that still plagues most AI search setups.

Prebuilt agent presets

OpenJarvis ships with several ready-to-run agent presets, each bundling the engines and tools needed for a specific task:

• Morning briefing pulls from a user's calendar, email, and the day's news to generate a digest. On macOS, it connects to Google Drive and local mail.
• Deep research answers complex questions by searching across the web and local documents, returning results with citations.
• Local coding assistant writes and runs Python scripts on the user's machine to complete tasks.

Each preset initializes with a simple command and can be customized through the framework's plugin system. For developers who have grown used to cloud-based coding agents like Cursor's enterprise-leading AI coding agents, the local approach adds a layer of control, but also a layer of responsibility for managing the underlying stack.

Competing with cloud giants

OpenJarvis enters a field dominated by cloud-backed offerings from OpenAI, Google, and Microsoft, where local inference has been treated as a secondary feature at best. The framework's emphasis on transparency and energy tracking appeals to privacy-conscious users, developers who work offline, and organizations whose compliance rules forbid sending data to external servers. The question is whether the efficiency gains can keep pace with the ecosystem that the cloud giants are building, a question that recent models like Gemma 4's rewrite of open-weight efficiency standards suggest is increasingly competitive.

The practical usefulness of local agents still depends on the user's hardware. Running a 35-billion-parameter model locally requires ample RAM and a capable GPU. The project acknowledges this. The startup experience is smoother on high-end machines, but the modular design lets users swap smaller models for daily tasks and pull in larger ones only when needed, an approach that mirrors the load-adaptive strategies emerging in the verification horizon for coding agents.

OpenJarvis is available now under an open-source license. The team at Stanford expects to release additional presets and model integrations over the coming months.