Stanford's OpenJarvis Brings a Personal AI Agent to Your Own Hardware via Ollama
A typical personal AI assistant sends every request, email and calendar entry to someone else's cloud server. OpenJarvis, an open framework from Stanford labs, runs the agent on your own computer, with the cloud as an option rather than a requirement. It now works through Ollama, a tool for running neural networks locally, and installs with one command. That makes it possible to set up a private assistant for mail, documents and code without an external API.
AI-processed from Ollama Blog; edited by Hamidun News
Imagine a personal AI assistant that has access to your email, calendar and work documents. In most cases, every question you ask and every email it reads goes to a cloud server you don't control, and you pay for that with money and privacy. Meanwhile, local models already handle most everyday tasks well: answering questions, summarizing text, drafting code. The Ollama team, which maintains a popular tool for running neural networks on your own computer, has announced that the open OpenJarvis framework can now run on top of it. The idea is simple: everything runs on your machine first, and the cloud connects only if you choose it.
What OpenJarvis is and where it came from
OpenJarvis is an open framework for building personal AI agents that run on your own hardware. It is developed by Stanford labs as part of the research program "Intelligence Per Watt", meaning "intelligence per watt". The point of the program is to measure not only the quality of answers but also how much energy, money and time each answer takes.
That is why OpenJarvis enables local models by default, and the cloud remains optional. The framework tracks energy, cost and latency alongside accuracy. If you are choosing between your own graphics card and a paid API, this helps: you see the price of each answer, not just its quality.
How to run it and what's inside
Installation takes one line. On macOS and Linux, you run the install script, which sets up the environment and finds an existing Ollama installation. On Windows, the same command runs in a separate environment; details are in the project's instructions. The script installs a starter model, so you can start working right away.
If you need a different model, you can pick any model available in Ollama. For example, jarvis model pull qwen3.5:35b, and then jarvis ask -m qwen3.5:35b "your query". To avoid specifying the model every time, you set it in the config: default_model = "qwen3.5:35b" and preferred_engine = "ollama".
The main convenience is the ready-made presets. Each one assembles an agent with the required engines and tools:
- morning-digest-mac builds a morning summary from your calendar, email and the day's news. Judging by the name, the preset is designed for macOS.
- deep-research searches the web and your local documents and answers with citations. You first index a folder of files with
jarvis memory index ./docs/, after which you can ask something like "Summarize all emails about Project X". - code-assistant writes and runs Python code directly on your machine to complete a task.
Any preset launches with jarvis init --preset <name>, with no manual assembly of the toolchain.
What this means in practice
For a developer, OpenJarvis is a ready-made starting point that doesn't need to be built from scratch: the agent, document memory, tools and engine selection are already wired together. For a product manager or founder, the more important point is different. Privacy here follows from how the system is designed: your mail and internal documents stay on the computer unless you enable the cloud yourself. This addresses many concerns for teams that cannot send data to external services.
There are limitations too. The announcement contains no benchmarks: Ollama does not provide figures for accuracy, energy use or speed on specific graphics cards and laptops. The qwen3.5:35b example shows how to connect a model, but it does not promise that the model will run comfortably on any hardware. A model with tens of billions of parameters requires a substantial amount of memory, so on a weak laptop you will need to choose something smaller. In addition, an agent that reads your email and runs code on your machine gets broad permissions. Before using it, check which access you are granting it.
If you have long wanted to try a personal agent without a subscription and without sending data out, this is one of the simplest ways to do it. You need only an installed Ollama, one command and a preset choice. And the built-in tracking of energy, cost and latency will help you fairly compare local runs with the cloud on your own tasks.
- Nov 21, 2026NAND flash for SSDs drops 19–32% in spot deals — drives could get cheaper
- Nov 21, 2026Unsloth Updates Compressed Qwen3.8-27B: Accuracy Up 5–10% at the Same Size
- Nov 21, 2026AI Writes Vulnerable Code in 44% of Tests: Positive Technologies Gives Agents a Scanner
- Nov 21, 2026Sber Releases a 14,990-Ruble Speaker with GigaChat That Understands Voice Without Internet
Key Questions Answered
What OpenJarvis is and where it came from?
OpenJarvis is an open framework for building personal AI agents that run on your own hardware. It is developed by Stanford labs as part of the research program "Intelligence Per Watt", meaning "intelligence per watt". The point of the program is to measure not only the quality of answers but also how much energy, money and time each answer takes. That is why OpenJarvis enables local models by default, and the cloud remains optional. The framework tracks energy, cost and latency alongside accuracy.
How to run it and what's inside?
Installation takes one line. On macOS and Linux, you run the install script, which sets up the environment and finds an existing Ollama installation. On Windows, the same command runs in a separate environment; details are in the project's instructions. The script installs a starter model, so you can start working right away. If you need a different model, you can pick any model available in Ollama. For example, jarvis model pull qwen3.5:35b, and then jarvis ask -m qwen3.5:35b "your query".
What this means in practice?
For a developer, OpenJarvis is a ready-made starting point that doesn't need to be built from scratch: the agent, document memory, tools and engine selection are already wired together. For a product manager or founder, the more important point is different. Privacy here follows from how the system is designed: your mail and internal documents stay on the computer unless you enable the cloud yourself. This addresses many concerns for teams that cannot send data to external services. There are limitations too.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.