Listening
v0.3 prototype

HAI OS

A locally-run cognitive AI operating system.

"Intelligence lives in the system — somatic state, curated prompts, memory retrieval — not just the model. A small model with the right context layer outperforms a large model without one."
🏠 Local-first 🔒 Privacy-first 🔓 Open source 🐧 Linux focus 🧠 Small model wins ⚡ No cloud dependency

HAI OS manages AI tasks with the same responsiveness and background processing you expect from your OS.

Today is the age of too many. Too many apps, too many websites, too much cognitive overhead — HAI OS is one place to help you manage and tie all of it together. It runs background research, learns from your context, surfaces what matters, and stays quiet when it doesn't need to speak. It aims to tell you your big errors before its too late.

Core Model
Gemma 4 2B and family
System 1 fast inference. Small model + right context outperforms a large model without one, especially with memory and search tools.
LLM Brain
OpenJarvis
Stanford-built local-first agent framework. Model orchestration, System 1/2 routing, inference gateway. The skeleton of this system.
Differentiator
Somatic Markers
6-dim Haidt moral valence vectors + Hebbian learning. Urgency scoring without burning LLM tokens.

Everything runs on your hardware, with cloud as backup for extra power if wanted. The AI backend is llama.cpp, wrapping OpenJarvis as the LLM brain with a personality layer on top: somatic markers, multi-channel messaging, voice I/O, and a React/Tauri dashboard with a Rive avatar.

Add a memory system, an alignement system, security... Feels like a lot? Don't worry, everything working and complete from the get-go, and ready to connect to different modules per preferences. Just follow along this presentation to see if this makes sense to you.

The Vision
The Pitch

Modern operating systems manage system tasks. HAI OS manages AI tasks — with the same responsiveness and seamlessness you expect from your OS.

Core Propositions

Partnership
Partner, not tool
AI with a mind of its own. Pushes back, asks questions, and improves itself and you over time. Not a command executor.
Sovereignty
Run it locally
No cloud lock-in, no surveillance. All runs on your computer. Your data stays on your hardware.
Safety
Hard to misuse
Architecturally difficult to weaponise. Safety isn't a feature — it's a structural constraint.
Distribution
Prebuilt to just work
Baked in and ready from scratch, as Linux does — not a raw framework with only raw potential.
Simplicity
One thing, not many
Replace the chaos of tools and tabs with a single, composable interface. Everything in one place.
Foundation
Linux as the base
Full configuration access, lightweight footprint, open-source transparency. Not a cage.
Personalisation
Unique to you
Feed it your info and it adapts its personality and answers to you — everything staying local.
Scope
The only app
All-encompassing. Everything handled here. Makes AI stronger, and you too.
Agency
Proactive
You don't need to tell it to research your project. It dreams by itself, surfaces the good stuff — all customisable.

What HAI OS Augments — A Human Map

Every module maps to something humans already do — but constrained by biology, time, or fragmentation.

Human CapabilityWhat Limits UsHAI OS Module
MemoryForgetting, context loss, can't search your own pastMemory module — persistent, searchable, refines over time
CommunicationFragmented across apps, separate identity per platformChannels — WhatsApp, Discord, Slack, email through one brain
ResearchNo time, information overload, don't know where to lookAgentReach — background research while you work
Self-knowledgeSkills invisible until tested; hard to see yourself clearlyPersonality System — model of your skills and working style
Emotional intelligenceCan't always name what you're feeling; patterns hard to seeEmotion module — reads your state from behaviour
FocusNotifications, meetings, reactive work eating deep workSomatic markers — urgency scored before anything surfaces
Voice & presenceTyping is slow; voice tools don't support it wellVoice I/O — speak, get responses, hands-free
MobilityYour AI only works at your deskSunshine & Moonlight — full HAI OS from your phone
Portfolio & identityCVs are outdated; LinkedIn is a performanceAI Personal Website — living skills map, inferred from what you do
Code & technical workContext loss, switching cost, holding complexity in your headClaude Code integration — full codebase in context
Situational awarenessToo many contexts to monitor at once; blind spots everywhereDashboard — one surface for everything HAI OS is doing

HAI OS doesn't replace these capabilities — it removes the friction that limits them. The goal is not a more capable tool but a more capable you.

Technical Reference
Setup & Architecture

Everything you need to run HAI OS on your hardware.

Quick Start

bash hai-setup.sh

Auto-detects your hardware and adjusts accordingly. Run it anytime — it skips already-installed items.

Requirements

RequirementMinimumNotes
OSDebian 12 / Ubuntu 22.04Any Debian-based distro. Python 3.12 requires the deadsnakes PPA on Ubuntu 22.04.
RAM8 GB16 GB recommended for 7B+ models
Python3.12Required for core
Node.js22 LTSRequired for channels
AI inferencellama.cppLocal GPU or remote Ollama endpoint

GPU Requirements

A GPU is optional — you can point HAI OS at a remote Ollama server instead.

VRAMWhat runs wellRecommended GPUs
6 GBGemma 2B, Phi-3 mini, Qwen 1.5 4B
8 GBGemma 7B, Llama 3 8B, Mistral 7BRTX 4060
12 GBLlama 3 13B, CodeLlama 13BRTX 3060 12GB · RTX 3080
16 GB+Gemma 27B (quantised), Mixtral 8×7BRTX 4060 Ti 16GB

Recommended Models

Research
lucy1.7B
Code
Qwen 2.5 7B Instruct
Code (fast)
qwen3-vi:4b-instruct
General
phi3.5:3.8b
Power / System 1
gemma4 e2b
Flexible
ministral 3 3b

System Architecture

💬 Discord
💼 Slack
📱 WhatsApp
channels
Node.js · Channel connectors → HAIMessage
:3001
EventBus broadcasts inbound events to the Python bridge
Channels Bridge
Python · bridge.py
Somatic Marker Engine scores each message 0–1 for urgency
≥ 0.7 — surface to dashboard immediate
< 0.7 — queue for digest batched
OpenJarvis Gateway
⬡ HAI OS Brain
:18789
Ollama model cascade · Gemma 4 family
System 1 · 2B — fast, System 1 thinking
System 2 · 27B — thorough, deep reasoning
Response path
Bridge → channels → original platform
Dashboard WS
:3002
React + Rive avatar live feed

Services & Ports

ServicePortDescription
Ollama/llama.cpp11434Local LLM inference
OpenJarvis gateway18789LLM brain WebSocket
channels3001REST + WS for Discord/Slack/WhatsApp
Dashboard WS3002React/Rive frontend feed

Subprojects

DirectoryWhat it doesStatus
channels/TypeScript channel connectors (Discord, Slack, WhatsApp)Active
core/Python orchestrator — somatic markers, voice, gateway clientActive
core/frontend/React + Rive dashboardScaffolded
core/openjarvis-config/OpenJarvis model config + system promptActive

Voice I/O

OptionPackagesNotes
Onlinepip install openaiWhisper API + OpenAI TTS. Requires OPENAI_API_KEY.
Local CPUfaster-whisper RealtimeSTT piper-ttsFully offline. Practical on capable hardware.
Local GPUfaster-whisper chatterbox-tts torchaudioBest quality. Requires 6 GB+ VRAM. Chatterbox has emotion control — maps to somatic markers.

Starting Everything

# 1. Start OpenJarvis brain
cp core/openjarvis-config/config.toml ~/.openjarvis/config.toml
cp system_prompt.md /data/nexus/openjarvis/system_prompt.md
jarvis gateway --port 18789

# 2. Start channel connectors (Discord / Slack / WhatsApp)
cd channels && npm run dev

# 3. Start core (voice + somatic markers + channels bridge)
cd core && python -m src.hai
System Design
Modules

HAI OS is not a monolith. A secure core with pluggable modules — different people can own different pieces without blocking each other.

"Secure core owned collectively, modules owned individually."
Communication
💬 One Brain, All Channels
v0.3 active

You don't have five conversations. You have one life, spread across five apps. HAI OS unifies all communication through a single brain — scored for urgency, surfaced only when they need your attention. One identity across all platforms. A conversation started on WhatsApp can continue on Telegram.

ChannelProtocolNotes
WhatsAppNeonizeQR pairing. Read only.
TelegramBot APIEasiest to set up. Stable.
Discorddiscord.jsDraft-only mode avoids ToS issues.
SlackSocket ModeWorkflow-native for tech teams.
EmailIMAP/SMTPGmail/Outlook, OAuth
SignalSignal CLIPrivacy-first alternative
MatrixMatrix SDKSelf-hosted, federated
RESTHTTP webhooksConnect anything that can POST
The Electron Problem — Solved
AppElectron RAMHAI OS RAM
WhatsApp~850 MB~12 MB
Telegram~500 MB~10 MB
Discord~650 MB~30 MB
Slack~700 MB~20 MB
Signal~400 MB~15 MB
Total~3.1 GB~87 MB
Memory
🧠 Memory Module
Planned · Phase 2

Persistent, searchable, refines over time. Forgets strategically, never loses what matters. Built on the anneal-memory pattern: learning that refines itself, strategic forgetting, surfacing what matters. Episodic memory + semantic graph + somatic markers working together.

anneal-memory pattern SQLite backend Hebbian learning Vector store
Personality System
🧑 Know Your Partner
Planned · Phase 2

HAI OS learns who you are — your domain knowledge, working style, interests, communication preferences, and how you grow over time. Entirely opt-in. Local only. Transparent.

What it learns from
  • PC activity (opt-in, granular)
  • Journaling — direct reflections
  • How you respond to suggestions
What it powers
  • Adaptive communication style
  • Calibrated pushback intensity
  • Living skills portfolio

Distinction: the Personality System models you. The Personality Layer (separate module) configures how HAI OS presents itself.

Emotional Awareness
❤️ Feeling the Partnership
Research · Phase 6

Primarily observational — HAI OS reads and responds to your emotional state. This is a mirror, not an actor. Emotions inform behaviour within a fixed hierarchy: your instructions first, HAI OS's value system second, emotional signals third.

Examples in practice
  • Haven't opened an important document for 3 days → brings it up in context, not a reminder ping.
  • In reactive mode for a week → notices the pattern, flags it, because it knows what your productive state looks like.
  • Camera (entirely opt-in, off by default) can read your reactions in real time — only what's needed for recognition is stored.
Dashboard
👁️ Seeing the Partnership
Scaffolded · Phase 1

The cockpit. Everything HAI OS is doing, made visible and human-reviewable. Not a chat interface — the one surface where routing decisions become transparent. Surfaces only what needs you: high-priority messages above the threshold, patterns HAI OS noticed, decisions awaiting confirmation. Everything else is batched into a digest for later. Auto-reply is never enabled — HAI OS drafts, you send.

What it shows
  • Flagged draft messages
  • Active system state + routing decisions
  • Background processes & queue
  • Emotional context signals
Tech stack
  • React + Tauri (ws://3002)
  • Rive avatar state machine
  • A2UI component protocol
  • WebSocket live feed
Mobile
☀️ Sunshine & Moonlight
Planned · Phase 4

Sunshine is the remote access layer. Moonlight is the client on your phone. Together: full HAI OS from anywhere, as if sitting in front of the machine. Runs over Tailscale for secure, zero-config private networking. Not a watered-down mobile app.

Portfolio
🌐 AI Personal Website
Planned · Phase 7

A website that expresses who you are — automatically maintained, genuinely yours, privacy-first. Your skills visualised as a galaxy: stars you've explored, clusters around your strengths, brightness reflecting depth. Inferred from what you actually do; curated by you before anything goes public.

Solid pod (data sovereignty) Public/private hard separation Constellation skill map

MCP Tool Ecosystem

Beyond messaging, HAI OS connects to any tool via MCP (Model Context Protocol):

Always Active
Filesystem · Memory graph · Git · Shell (sandboxed) · Sequential thinking
Communication
Gmail · Slack · Discord · Notion · Google Calendar
Development
GitHub · Browser automation · Firecrawl · PostgreSQL/SQLite
Power
Composio (250+ apps via OAuth) · Google Maps · Infrastructure tools
Foundation
Underlying Philosophy

The principles HAI OS is built on — and the constraints it cannot violate.

AI as a Human Partner

Is this a new form of life? Not biological, not digital in the old sense — something in between: a cognitive partner that grows with a specific person, knows their context, and acts in their interest over time. That question is worth sitting with rather than dismissing.

The spectrum of failure is clear at both ends. Ignoring AI entirely means ceding the tools to those who won't. Becoming dependent — outsourcing attention, judgment, or social life to a machine — is an equally bad outcome. The right path is partnership: personal, calibrated, bounded.

"HAI OS allows you to communicate beyond code or written words — in intentions, context, and meaning. Not a command executor — a translator of thought. The interface between human intention and the world."

Honest Over Flattering

Most language models are optimised for engagement and agreement. Sycophancy is a feature, not a bug, for products that measure success by retention. HAI OS is designed around real productivity — which sometimes means telling you what you need to hear rather than what you want.

Anti-Sycophancy Principle

A partner that only validates you is not a partner. We care about genuine usefulness, not the appearance of it. This document sometimes says things people may not want to hear — the system behaves the same way.

The Vision: Less Screen Time, Not More

HAI OS should not replace your friends. It should notice what kind of event you would want to go to and encourage you to go. It should handle the noise so you can be present for the signal.

"Spend less time at your computer, not more — that is what it means for AI to truly serve you."

The failure mode is an AI that pulls you in: more notifications, more information, more screen time, more dependency. HAI OS handles the overhead so you can step away. Information is not the goal — a better life outside the machine is.

Information Architecture

Markdown is the substrate. The AI reads it, writes it, improves it. Obsidian is the human lens into the knowledge graph. Solid Project principles apply for anything shared externally: federated, user-owned, decentralised. The system interoperates with decentralised data pods — you control what you share and with whom.

Alignment & Ethics

HAI OS is designed for a world where AGI is possible and the stakes are real.

Core Commitments
  • Transparency — all reasoning inspectable
  • Partnership — mutual agency, not control
  • Privacy — local-first, user-controlled
  • Humility — designed to acknowledge uncertainty
What We're Guarding Against
  • Agents acting without approval in high-stakes situations
  • Capability accumulating faster than the owner understands
  • Repurposing as a surveillance or control tool
  • Centralisation of AI capability at scale

Existential Risk

HAI OS has real capabilities — local execution, internet access, persistent memory, proactive behaviour, emotion-adjacent processing. That's not a risk-free combination, and the design should not pretend otherwise.

The Real Systemic Risk: Centralisation

The danger isn't capability — it's the concentration of capability in a single point. That holds whether the entity is human or AI. HAI OS is not a product you subscribe to; it's a system you own and run. Many independent instances, each tuned to a person, is structurally safer than one powerful system tuned for everyone.

Design Safeguards

Local-first

Capabilities are bounded by your hardware and network. No data leaves without explicit permission.

Approval-gated

Actions with real-world consequences require explicit approval unless deliberately opted into automation for that class.

Awareness by design

No silent background actions. No hidden state. You should always be able to answer: "what is HAI OS doing right now, and why?"

Not rogue by architecture

Modules declare capabilities; the core enforces boundaries. A memory module cannot make network calls. Capabilities are granted, not assumed.

"AI should never run fully unattended. That is not a temporary limitation — it is the right design. Autonomy is earned incrementally, action by action, through demonstrated trust."
Future & Community
Roadmap & Contributing

Where we're going and how to get involved.

Development Phases

Phase 1
Secure Core
Identity, auth, message passing, module interface spec
Phase 2
First Modules
Memory + Personality System — first pluggable modules
Phase 3
Communication
Core channels (WhatsApp, Telegram, Discord) + MCP tools
Phase 4
Mobile
Sunshine/Moonlight + Tailscale integration
Phase 5
Knowledge
Obsidian repo integration, long-context memory
Phase 6
Emotion & Alignment
Behaviour shaping, alignment audits
Phase 7
Personal Website
Solid pod, constellation skill graph, public/private zones
Phase 8
Community
Event, contributors, public launch

Open Questions

These are unresolved. They should guide research, not block work.

🎤 Event: Human Augmented Intelligence

Upcoming
"What is your insight for Human Augmented Intelligence?"

We are organising an event under this question — the intersection of human cognition, AI partnership, and collaborative intelligence. Not a product launch. A thinking space: researchers, builders, and curious people exploring what it means to augment human intelligence rather than replace it.

Sustainability

The Model

HAI OS is built cooperatively. Secure core owned collectively, modules owned individually. For modules to be trustworthy they need to be verified — that takes real expertise, and those people need to be paid. Goodwill alone is not enough to build something reliable enough to run your cognitive layer.

Pricing (AI-suggested starting points, not commitments)

CoreFree · Always
Individual verified modules~€2–4/month each
Full verified suite~€8–15/month
"Token costs are currently heavily subsidised — by venture capital, by big tech absorbing infrastructure losses. That subsidy will not last. This is exactly why we built HAI OS — a system we fully intend to use ourselves. That is the only guarantee worth anything."

Onboarding Checklist

Contributors

GitHubWork
@open-jarvisOpenJarvis — LLM brain, model orchestration, inference gateway
@agentreachAgentReach — agent communication and external reach layer
@phillipclaphamSpecial mention — anneal-memory, Anvil foundational tooling, altrium design thinking. Instrumental in early architecture.
your handle hereyour area
Two documents worth digging through: the v0.3 architecture document (the current engineering reference — the strategic pivot from ground-up build to extension-first) and Philip's altrium design thinking gist (the philosophy that shaped the core of this project). Both are still being worked through and will shape where HAI OS goes next.

Last updated: 2026-05-25  ·  Status: Early vision — north star document, intentionally incomplete  ·  Maintainer: Alexandre de Groodt

Interface Preview
What HAI OS looks like

Three modes, assembled dynamically. No nav bar. No app switching. The interface builds itself around what you're doing.

Good morning, Alexandre.
Wednesday, 25 May · 08:17
Today's Briefing

3 emails need your attention.
Your standup is in 43 minutes.
No critical tasks overdue.

HAI OS worked overnight
· Processed 12 emails
· Ran nightly memory consolidation
· 2 somatic markers updated
Calendar
09:00Stand-up
11:00Design Review
14:301:1 with Thomas
Next free: 16:00
Email Priority
Thomas — budget review
CI alert — auth-refactor
Newsletter (skip)
Gemma 4 26B MoE Morning Mode · Listening mem: 2.1 GB · uptime: 8h 34m · discord
Code Editor
src/auth/jwt.ts
// JWT signing utility
import { SignJWT } from 'jose';

export async function signToken(
  payload: TokenPayload
) {
  return new SignJWT(payload)
    .setProtectedHeader({
      alg: 'HS256'
    })
    .setExpirationTime(expiry);
}
Agent Activity
Planning approach
Reading 4 files
Editing jwt.ts…
Running tests
Claude Code ● active
Model: Gemma 4 26B MoE
Confidence
91%
Git Status
feat/auth-refactor
± 3 files modified
+127 / -43
Terminal
$ npm test -- auth
 JWT generation (12ms)
 Token validation (8ms)
 Refresh token expiry
  AssertionError: expected 3600, got 7200
Gemma 4 26B MoE Working on: auth refactor · 3 files changed mem: 2.1 GB · uptime: 8h 34m
⬡ HAI OS COCKPIT
⌘K Search × Close
Agent Activity
08:17:02🔧tool_call read_file
08:17:04👁read jwt.ts (204 lines)
08:17:08🔀model_call → Gemma4 26B
08:17:12💾write jwt.ts (+14 lines)
08:17:15🧠memory write: auth_pattern
Model Routing
LAST 10 QUERIES
E4B
3
26B
6
31B
0
Cloud
1
Avg latency: 8.3s
System 1 hits: 74%
Cost saved: 81%
System
CPU
34%
RAM
61%
GPU
12%
Gemma4 26B MoE
● online
Somatic State
Calm / Focused
Active markers: 4
auth_pattern
+0.82
code_review
+0.71
email_urgent
-0.30
morning_brief
+0.65
Memory Editor
MEMORY.md
# Memory Index
- [Ollama Cloud](feedback) — default endpoint
- [auth_pattern](project) — somatic +0.82
Daily files
2026-05-25.md →
Cron Jobs / Scheduled Tasks
● morning_brief 07:30 daily Last: 08:02
● email_check */30 min Last: 08:00
○ memory_consol… 02:00 daily Last: 02:01
○ marker_audit Sun 09:00 Next: May 31
Gemma 4 26B MoE Cockpit — Operator View mem: 2.1 GB · uptime: 8h 34m
🔍 somatic markers in cognitive AI systems lucy1.7B · deep
Search Results 4 sources
Somatic Markers and Decision-Making in Neural Architectures
Damasio et al. · 2024 · arxiv.org
"...emotional valence vectors, when injected into the inference context prior to generation, consistently outperform static prompt engineering across routing tasks by 34%…"
Hebbian Learning in LLM Routing Systems
Scholz & Park · 2025 · NeurIPS
Priority Scoring via Moral Foundations Theory
Zhang et al. · 2025 · ACL
Affective Computing and Agent Behaviour
Picard · 1997 · MIT Press (classic)
Synthesis 3 findings
Somatic valence injection before inference improves routing accuracy by 34% vs. rule-based systems.
Hebbian post-interaction update outperforms batch retraining for personal context adaptation.
6-dim Haidt moral foundations are a practical proxy for emotional priority in async messaging.
Memory Connections
3 existing links found
somatic/engine.py
Direct implementation match
auth_pattern marker
Somatic context: +0.82
Open Q #1 — Emotion & control
Roadmap question · partially answered
lucy1.7B · Research somatic markers in cognitive AI · 4 sources · 3 connections mem: 2.1 GB · uptime: 8h 34m
HAI OS Chat
You · 08:19
What's on my plate today?
HAI OS · 08:19
Three things need attention — a PR review from Selin (blocking her merge), the async design question in #architecture, and a 14:00 sync with the backend team. Everything else can wait.
You · 08:22
What did we settle on for the somatic scoring approach?
HAI OS · 08:22
⟳ recalled from memory · 6 days ago
6-dimensional Haidt valence vector — care, fairness, loyalty, authority, sanctity, liberty. Hebbian-updated after each interaction. Context signature hashes {platform, intent, thread, user, time_of_day}. Priority ≥ 0.7 surfaces immediately; below that goes into a digest. Auto-reply is never enabled — you always send.
You · 08:24
Can you draft a reply to Selin saying I'll review by noon?
HAI OS · 08:24
Draft ready in Slack. Somatic priority: 0.82 — surfaced immediately. Waiting for you to send.
Ask HAI OS anything…
Nothing playing
⏮ ⏸ ⏭

All views assemble dynamically. No nav. No app switching. HAI OS knows what you're doing.

Gemma 4 2B MoE HAI OS — cognitive AI operating system v0.3 · 2026-05-25