Personal AI Operating Layer

Zora, always
running.

Zora OS is a persistent AI that lives on your hardware. It remembers your context, connects your channels, and acts on your behalf — running locally on Apple Silicon or Linux.

$ curl -fsSL https://zoraos.ai/install.sh | bash
Apache 2.0 / GitHub / Apple Silicon & Linux

Not a chatbot. Not a copilot.
A persistent digital extension of you.

Zora maintains context across conversations, days, and weeks. She triages messages, preps for meetings, reflects on what she's learned — whether you're talking to her or not.

Capabilities

What the system does.

01

Local brain

Direct Metal inference via MLX on Apple Silicon. A fine-tuned 4B orchestration model at ~90 tok/s. No API calls. No cloud dependency.

02

Persistent memory

Vector, full-text search, and knowledge graph memory that compounds over time. Never starts from zero.

03

Multi-channel

WhatsApp, Telegram, Teams, email, Slack, Discord — one nervous system. Triage, route, and respond across all channels.

04

148 native tools

Smart handoff routes tasks automatically. Browser, code, search, files, infrastructure. The right tool, chosen for you.

05

Distributed compute

Any Mac joins the cluster. The orchestrator routes fast; workers run heavier models. 70B inference at ~90 tok/s via Metal.

06

Cognition loop

A background loop runs whether you're watching or not. Morning briefings, midday sensing, overnight reflection and consolidation.

Product

Built to be used, not just demoed.

Dashboard
Zora OS Dashboard
Architecture

Local-first. Private by default.

┌──────────────────────────────────┐ │ Dashboard │ │ Chat · Office · VR · Monitor │ └──────────────┬───────────────────┘ │ ┌──────────────▼───────────────────┐ │ FastAPI Orchestrator │ │ │ │ Cognition Agent Runner │ │ Memory Channels │ │ Trust Delegate │ └──────┬──────────────┬────────────┘ │ │ ┌──────▼─────┐ ┌─────▼────────────┐ │ Mac Mini │ │ Worker Nodes │ │ Zora-4B │ │ 70B via Metal │ │ ~7GB total │ │ ~90 tok/s │ └────────────┘ └──────────────────┘
01
Direct Metal inference

MLX runs the model directly on Apple's GPU. A custom TurboQuant Metal kernel compresses the KV cache from 4GB to 1.5GB at 32K context.

02
One nervous system

Every message, event, and integration flows through a universal event bus. Auto-classification, trust policy, and action — or approval — in seconds.

03
Privacy at the architecture level

Data stays on your machine. Secrets in macOS Keychain. Search via local SearXNG. A sanitizer strips sensitive content before anything leaves your device.

Memory & Reflection

She remembers. She reflects.

Daily Diary 2026-04-09
08:32
Reviewed overnight messages. Marcus sent project timeline update on WhatsApp. Sarah's meeting moved to 3pm. Drafted briefing.
14:15
Noticed recurring pattern: authentication questions from the team this week. Created consolidated reference from last three conversations.
23:00
Merged 12 short-term memories into knowledge graph. Updated relationship health scores. Prepared tomorrow's morning briefing.
People Intelligence continuous

Learns facts from every conversation. What someone does, where they live, what they care about. Before a meeting, she generates talking points specific to that person.

Tracks relationship health and nudges you when someone is going cold. Learns your outgoing tone — formality, greetings, sign-off style — and drafts replies accordingly.

Always draft-and-approve. Never auto-sends. Trust boundaries are visible and reviewable.

Get Started

Run Zora on your hardware.

Self-hosted. Fully private. No account required.

$ curl -fsSL https://zoraos.ai/install.sh | bash
Source Docs Community
Apple Silicon or Linux · 16GB+ RAM · Python 3.10+
Copied