OpenJarvis: Stanford's Local-First Personal AI Agent Framework
The pitch for personal AI agents has always had a catch: your “personal” AI runs on someone else’s servers. OpenJarvis, from Stanford’s Scaling Intelligence lab, is built around a different premise — local first, cloud only when necessary.
What OpenJarvis Is
OpenJarvis is a framework for building personal AI agents that run on your own hardware. It provides shared primitives for on-device agents, evaluations that treat energy, latency, and cost as first-class constraints (not just accuracy), and a learning loop that improves models using your local trace data.
The project is grounded in Stanford’s Intelligence Per Watt research, which found that local language models already handle 88.7% of single-turn chat and reasoning queries, with efficiency improving 5.3× from 2023 to 2025. The models are ready. The software stack was what was missing.
Getting Started
Installation is a one-liner:
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash
jarvisAbout 3 minutes on broadband. It handles uv, Python venv, Ollama, and a starter model automatically. Then jarvis doctor shows you what’s running.
Starter Presets
OpenJarvis ships with preset configurations for common use cases:
| Preset | What It Does |
|---|---|
morning-digest-mac | Spoken briefing from email, calendar, health, news |
morning-digest-linux | Same with vLLM support for GPU servers |
morning-digest-minimal | Gmail + Calendar only, runs anywhere |
deep-research | Multi-hop research assistant |
code-assistant | Coding helper |
scheduled-monitor | Periodic monitoring tasks |
chat-simple | Basic conversational agent |
Switch presets with:
uv run jarvis init --preset morning-digest-macWhy Local-First Matters
The case for running AI locally keeps getting stronger:
- Privacy — your emails, calendar, health data, and browsing history stay on your machine
- Cost — no API bills for daily briefings and routine queries
- Latency — no round-trips to cloud servers for simple tasks
- Reliability — works offline, no service outages
- Customization — fine-tune on your own data without uploading it anywhere
The 88.7% coverage stat is key. Most personal AI tasks (summarize this email, what’s on my calendar, explain this code) don’t need frontier models. A well-tuned 8B model on an M-series Mac handles them fine.
How It Compares to the Field
The personal AI agent space has exploded in 2026. Here’s how OpenJarvis stacks up against the major players across the dimensions that actually matter:
Feature Comparison
| Feature | OpenJarvis | OpenClaw | NemoClaw | HolaOS | Lore |
|---|---|---|---|---|---|
| Primary focus | Local-first agents | Multi-channel assistant | Enterprise security | Desktop agent computer | Local memory/recall |
| Default execution | Local | Cloud (local optional) | Cloud + sandboxed | Local (Electron) | Local |
| Model support | Ollama (local) | Claude, GPT, Ollama | Any + Nemotron | Any LLM provider | Ollama only |
| Messaging channels | None built-in | 8+ (Telegram, WhatsApp, Slack, Discord, Signal, iMessage…) | Inherits OpenClaw’s | None | None |
| Memory persistence | Local trace learning | Daily notes + MEMORY.md | OpenClaw + enterprise audit | Per-workspace AGENTS.md | LanceDB vector store |
| Skill/plugin system | Preset configs | 5,400+ community skills | Inherits OpenClaw’s | Apps + Skills markdown | Instruction memory |
| Security model | Local by default | User-managed | OpenShell sandbox + guardrails | Workspace isolation | Local by default |
| Hardware requirements | Consumer GPU/CPU | Any (API-driven) | Any (hardware agnostic) | macOS (Electron) | Any (Ollama) |
| Backing | Stanford research lab | Open-source community | NVIDIA | Independent | Independent |
| Best for | Research, efficiency-conscious users | Power users wanting integrations | Enterprise deployments | Visual desktop workflows | Personal knowledge base |
The Practical Differences
OpenJarvis vs OpenClaw — This is the most interesting comparison. OpenClaw is the Swiss Army knife: 8+ messaging channels, browser control, cron jobs, 5,400+ skills, shell access. It’s cloud-first (Claude/GPT by default) but supports local models via Ollama. OpenJarvis is the opposite philosophy — local-first with cloud as fallback. If you want an agent that connects to your Telegram, WhatsApp, and Slack simultaneously while managing your calendar, OpenClaw wins. If you want an agent that runs your morning briefing entirely on-device without API costs, OpenJarvis wins. They’re solving different problems despite both being “personal AI agents.”
OpenJarvis vs NemoClaw — NemoClaw is NVIDIA’s enterprise wrapper around OpenClaw, announced at GTC 2026. It adds OpenShell sandboxing, policy-based guardrails, and privacy routing. Where OpenJarvis optimizes for efficiency on personal hardware, NemoClaw optimizes for security and compliance in enterprise environments. Different target users entirely — researcher vs. CISO.
OpenJarvis vs HolaOS — HolaOS gives agents a full desktop environment (browser, files, apps) via Electron. It’s about the agent living on your desktop as a persistent collaborator. OpenJarvis is more task-oriented — run a preset, get a result. HolaOS is richer for visual/interactive work; OpenJarvis is leaner for automated routines.
OpenJarvis vs Lore — Lore is narrower but deeper on one thing: personal memory. Global hotkey capture, automatic classification, RAG retrieval over your thoughts. OpenJarvis is a broader agent framework; Lore is a second brain. They’re complementary — you could run both.
OpenJarvis vs the lightweight alternatives — Our personal AI agent comparison covers NanoClaw, ZeroClaw, PicoClaw, and Luna Agent. Most of these are minimal runtimes (<10MB RAM). OpenJarvis sits between these and OpenClaw in complexity — more opinionated than a bare runtime, less feature-packed than a full assistant platform.
When to Choose OpenJarvis
Pick OpenJarvis if:
- You want zero API costs for routine personal AI tasks
- You care about energy efficiency and model benchmarking on real hardware
- You want preset workflows (morning briefing, research, code assist) that just work
- You’re on Apple Silicon or a Linux GPU box and want to maximize local inference
- You’re a researcher interested in the efficiency frontier
Pick something else if:
- You need multi-channel messaging (→ OpenClaw)
- You need enterprise security and compliance (→ NemoClaw)
- You want a full desktop agent environment (→ HolaOS)
- You want the absolute smallest footprint (→ NanoClaw or ZeroClaw)
The Learning Loop
The most interesting architectural choice: OpenJarvis includes a learning loop that uses your local interaction traces to improve model performance over time. Your agent gets better at your tasks without sending data anywhere.
No other framework in this space does this. OpenClaw has persistent memory (daily notes, MEMORY.md), HolaOS has workspace memory, Lore has vector-stored recall — but none of them fine-tune the underlying model on your usage patterns. OpenJarvis treats your interaction history as training data, not just retrieval context.
This is the personal AI endgame — not just a model that runs locally, but one that adapts to you locally.
Platform Support
- macOS (Intel + Apple Silicon) — best experience on M-series chips
- Linux — full support, vLLM for GPU servers
- Windows via WSL2 or desktop binary
- Desktop app available for non-terminal users
The Bigger Picture
We’re watching the personal AI agent stack bifurcate: cloud-powered agents that do everything through APIs, and local-first agents that prioritize privacy and efficiency. The AI OS comparison we published in March mapped the cloud vs. local divide across OpenClaw, Jeriko, and Perplexity. OpenJarvis is the strongest entry yet in the local-first camp, backed by real research rather than just ideology.
With NVIDIA betting on the agent stack via NemoClaw, Stanford publishing efficiency research that validates local inference, and the lightweight agent ecosystem proving accessibility — the infrastructure for personal AI is maturing fast.
If Stanford’s efficiency numbers hold — and the trend suggests they will — the question isn’t whether local AI agents are viable. It’s when they become the default.