OpenJarvis v1.0: Running Personal AI Agents on Your Own Hardware — A Practical Guide

On May 16, 2026, OpenJarvis hit v1.0 — a milestone for anyone who believes AI agents should run on your hardware, not in someone else’s cloud. Born from Stanford’s Hazy Research lab and backed by a team led by Jon Saad-Falcon, Avanika Narayan, and Christopher Ré, OpenJarvis is now the most mature open-source framework for building personal AI agents that run entirely locally.

We’ve been testing it for two weeks on a Mac Mini M4 Pro. Here’s what it does, how to set it up, and whether it’s ready for your SME.

What OpenJarvis Actually Is

OpenJarvis is not a chatbot. It’s a composable framework for building on-device AI systems — agents that can schedule tasks, search the web, read your files, execute code, and improve themselves over time, all running on your Mac without sending data to any cloud service.

The framework is built around five composable primitives:

Layer	Purpose	What You Get
Intelligence	Model selection	Auto-selects the best model for your hardware, or you pick manually
Engine	Inference runtime	Ollama, vLLM, SGLang, llama.cpp, plus cloud APIs as fallback
Agents	Multi-step reasoning	8 built-in agent types for different workflows
Tools & Memory	Capabilities	Web search, calculator, file I/O, code execution, MCP servers, persistent state
Learning	Self-improvement	SFT, GRPO, DSPy prompt optimization, GEPA agent logic, spec search

The key differentiator from other agent frameworks: OpenJarvis treats energy, FLOPs, latency, and cost as first-class constraints alongside accuracy. This comes from their “Intelligence Per Watt” research, which found that local LLMs already handle 88.7% of single-turn chat/reasoning queries, with efficiency improving 5.3x from 2023 to 2025.

The 8 Built-in Agents

OpenJarvis ships with 8 agent types out of the box:

Agent	Type	What It Does	SME Use Case
morning_digest	Scheduled	Daily briefing from email/calendar/news	Executive morning briefing
deep_research	On-demand	Multi-hop research with citations	Market research, competitive analysis
monitor_operative	Continuous	Long-horizon monitoring with memory compression	Brand monitoring, price tracking
orchestrator	On-demand	Multi-turn reasoning with automatic tool selection	Complex project management
native_react	On-demand	Thought-Action-Observation loop	Step-by-step problem solving
operative	Continuous	Persistent autonomous agent with state	Customer service bot, data pipeline
native_openhands	On-demand	Python code execution	Data analysis, report generation
simple	On-demand	Single-turn chat, no tools	Quick Q&A

The morning_digest agent alone is worth the setup time. It compiles your calendar, emails, and relevant news into a 2-minute audio briefing every morning. The deep_research agent does multi-hop web research with citations — we’ve been using it for EU AI Act compliance research and it’s surprisingly good at finding regulatory documents.

Setup: From Zero to Running in Under 2 Minutes

The claim is “setup OpenJarvis in under two minutes with a single Ollama command.” We timed it.

Step 1: Install

# macOS / Linux / WSL2
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash

The installer handles: uv (Python package manager), Python venv, Ollama, a starter model (qwen3.5:2b), and auto-detects your hardware to recommend the best engine.

Step 2: Start Chatting

jarvis
# First run: downloads Qwen 3.5 2B (~1.5GB), starts Ollama if not running
# Subsequent runs: instant startup

jarvis init --preset morning-digest-mac
# Sets up the daily briefing agent with calendar + email + news

Step 3: Try Deep Research

jarvis ask "What are the latest EU AI Act enforcement actions in Spain?"
# Uses deep_research agent: searches, reads, synthesizes, cites sources

Our Timing Results

Step	Time	Notes
Install script	45s	Includes Ollama + model download
First jarvis run	12s	Model warm-up on M4 Pro
Subsequent runs	3s	Near-instant on M4 Pro
morning_digest preset	8s	Configuration only
deep_research query	25–60s	Depends on search depth

Verdict: The 2-minute claim is accurate on a Mac with Ollama already installed. From zero to first query is about 90 seconds. From zero with Ollama installed: 12 seconds.

Local-Cloud Collaboration: The Smart Routing Pattern

OpenJarvis ships with three hybrid patterns for when local models aren’t enough:

Pattern 1: Per-Query Routing (60–80% Cost Savings)

The complexity analyzer evaluates each query. Simple ones stay local; complex ones escalate to a cloud model:

"Summarize this email" → Local (Qwen 3.5 2B) → 0.001€, 0.3s
"Draft a legal contract" → Cloud (Claude Sonnet) → 0.02€, 5s
"Translate to Spanish" → Local (Qwen 3.5 2B) → 0.001€, 0.3s
"Write a compliance audit" → Cloud (Claude Sonnet) → 0.03€, 12s

For a typical SME workload, this routing saves 60–80% on AI costs while keeping response times under 5 seconds for 85% of queries.

Pattern 2: LLM-Guided Spec Search

A frontier model proposes edits to the agent’s configuration (prompts, tool selection, model choice). Only non-regressing edits are accepted. This is how OpenJarvis agents self-improve over time.

Pattern 3: Minions-Style Decompose-and-Execute

The frontier model plans, the local model executes subtasks in parallel, and the frontier model aggregates. This is the most compute-intensive pattern but handles complex multi-step workflows well.

Production Readiness: Our Assessment

After two weeks of daily use, here’s our honest assessment:

What Works Well

✅ Morning digest agent — Excellent for daily briefings. TTS is clear, research is relevant, email/calendar integration is seamless. ✅ Deep research agent — Multi-hop citations are accurate. We used it for EU AI Act research and got better results than manual web searching. ✅ Local-only mode — We ran for 5 days with zero cloud API calls. All queries handled by local Ollama models. ✅ Hardware auto-detection — Detected M4 Pro, recommended Qwen 3.5 9B, set appropriate context window. No manual configuration needed. ✅ Energy telemetry — The 50ms resolution energy monitoring is unique and genuinely useful for cost calculations.

What Needs Work

⚠️ Learning loop — The SFT/GRPO self-improvement requires 100+ examples before it outperforms prompt engineering. Not practical for most SMEs yet. ⚠️ Desktop GUI — The desktop app (.dmg) works but feels alpha. Missing settings, no way to customize agent prompts from the GUI. ⚠️ Documentation — The GitHub README is good, but there’s no API reference. Python SDK docs are auto-generated and sparse. ⚠️ Error handling — When Ollama crashes or runs out of memory, OpenJarvis doesn’t always recover gracefully. Requires manual restart. ⚠️ No multi-user — Designed for single user. No authentication, no session isolation. Not suitable for team deployment yet.

What Doesn’t Work Yet

❌ Continuous monitoring agent (monitor_operative) — Memory compression still loses important context after 24+ hours. ❌ Plugin ecosystem — No third-party plugins or marketplace yet. ❌ Mobile — No iOS/Android app; remote access requires SSH tunnel.

Comparison: OpenJarvis vs. Alternatives

Feature	OpenJarvis	AutoGPT	CrewAI	LangGraph
Local-first	✅ Yes	❌ No	❌ No	❌ No
Ollama integration	✅ Native	🔶 Via API	🔶 Via API	🔶 Via API
Self-improvement	✅ SFT/GRPO	❌ No	❌ No	❌ No
Energy telemetry	✅ 50ms resolution	❌ No	❌ No	❌ No
Built-in agents	8 types	1 (general)	Role-based	Custom
Web search	✅ Built-in	🔶 Plugin	🔶 Plugin	🔶 Custom
Production ready	🔶 Beta v1.0	❌ Alpha	✅ Yes	✅ Yes
Team/multi-user	❌ No	❌ No	✅ Yes	✅ Yes
License	Apache 2.0	MIT	MIT	MIT

Our Recommendation for SMEs

For individual knowledge workers: OpenJarvis is ready. The morning digest and deep research agents alone justify the setup time. Install it on your Mac, run it local-only, and save €50–100/month on cloud AI subscriptions.

For teams: Not yet. Wait for v1.1 which promises multi-user support and a proper API. In the meantime, use OpenJarvis for individual productivity and CrewAI or LangGraph for team workflows.

For compliance-sensitive work: OpenJarvis is the best option for fully local AI agent workflows. No data leaves your machine in local-only mode. The energy telemetry helps you document compute costs for EU AI Act Article 53 reporting.

Sources

• OpenJarvis GitHub Repository
• OpenJarvis Documentation
• OpenJarvis v1.0.0 Release
• Intelligence Per Watt Paper (arXiv)
• Ollama MLX Benchmarks

---

Considering local AI agents for your team? Schedule a consultation — we’ll help you evaluate OpenJarvis against your specific requirements and set up a pilot deployment.