The chat-with-an-LLM era is over. What replaces it, for people who care about where their data lives, is finally arriving in a form that does not require a research grant to deploy. OpenJarvis v1.0 — released by Stanford’s Hazy Research and Scaling Intelligence labs, with native Ollama support — is the first credible open-source framework for personal AI agents that run on your own hardware. For UK operators who have watched the agent wave crash through 2025 and early 2026 from the side-lines, this is the moment to lean in.
What OpenJarvis actually is
OpenJarvis is not another chat client, and it is not a model. It is an agent framework — a layer that orchestrates local models, tools, and your personal data into something closer to a junior colleague than a search box. The thesis from Hazy Research is unflattering to most of the consumer AI industry: local models are already good enough for the day-to-day work of an individual operator, and shipping every prompt to a hyperscaler is a choice, not a necessity. Energy, cost, and latency are tracked alongside accuracy — a quiet admission that “intelligence per watt” is now a first-class metric, not a footnote.
The framework ships with three things that matter to a small-team operator:
- A clean install path that auto-detects an existing Ollama installation on macOS or Linux (or runs inside WSL2 on Windows), so the friction for the typical small-team setup is close to zero.
- A model-agnostic runtime that pulls any model Ollama serves — qwen3.5:35b is the example in the docs, but nothing in the framework cares which weights you point it at.
- Presets that bundle agents, engines, and tools into one-liners: morning briefing, deep research across local files, and a local code agent that writes and runs Python on your machine.
That last point is the part the press releases underplay. Presets are how frameworks either become infrastructure or stay curiosities.
The Ollama hookup — and why it matters
OpenJarvis is not built in a vacuum. It lands on top of the local-inference stack that has matured hard over the last 18 months. We have written before about Ollama v0.24 and its Apple Silicon gains and the head-to-head with LM Studio for small teams choosing a runtime. What OpenJarvis adds is the agent layer those runtimes never quite had — orchestration, memory, tool use — without reaching back to the cloud by default.
If you are already running local models for chat, OpenJarvis is the missing piece between “I have a model on my laptop” and “my laptop is doing agentic work for me every morning.”
The strategic logic is also worth naming. Stanford’s Hazy Research group is publishing open-source software that defaults to local execution in a year when every commercial lab is racing to make the cloud the only place you can run a serious agent. That is not a coincidence. The MCP protocol going stateless earlier this year, the rise of long-context open models like Llama 4 Scout, and the arrival of credible open-weight frontier models such as MiniMax M3 have all been pushing toward this moment. OpenJarvis is the integration layer that pulls the threads together.
What it does today, and what it does not
Realistic expectations matter. The morning-digest preset is genuinely useful: it pulls your calendar, email, and the day’s news into a single briefing, all on hardware you own. The deep-research preset, which indexes a local ./docs/ folder and answers questions with citations, is the kind of thing a sole-trader consultant will set up once and quietly depend on for years. The code-assistant preset is more aspirational — fine for one-off scripts, not yet a replacement for a proper IDE.
The framework is also refreshingly unopinionated about the model you bring. If you have been following the Qwen 3.6 conversation for 24GB cards, or have read our piece on running local AI on AMD hardware, you already know what is feasible on your own kit. OpenJarvis inherits all of that. It does not move the model-quality ceiling; it raises the productivity floor.
What to do this evening
If you are a UK operator with an existing Ollama setup, the deployment cost is an evening and a cup of tea: one curl command, one jarvis init --preset morning-digest-mac (or the deep-research equivalent), and an hour of fiddling with ~/.openjarvis/config.toml. If you do not have a local-inference setup yet, this is also a fair time to start — modest hardware will do, and the data never leaves your desk.
The verdict is straightforward. OpenJarvis v1.0 is the first time an open-source, local-first agent framework has arrived in a shape an operator — not a research lab — can actually use. It will not replace your CRM, your accountant, or your judgement. It will replace the part of your week spent copy-pasting between five browser tabs and one model. For a small UK firm, that is the most interesting kind of agent: the one you do not have to justify to procurement.
