Dina The Architecture of Agency

A personal AI agent completely loyal to you and only you.
Inspired by the novel UTOPAI (2012-2017).

Zero lock-in — your life is a file, not a hostage. Dina breaks you free from walled gardens.

Layers

Containers

Storage Tiers

Networking Options

Philosophy & Laws

Every design decision honours these four laws. They are not aspirational — they are engineering constraints enforced by architecture.

Dina isn't built to be your friend — she's built to help you find real ones. When you're lonely, she connects you to humans.

Core Principles

Dina does the drafting. You hold the keys. Agents can't send — only you approve.

The Trust Network

What if purchase decisions were based on real outcomes from real people? The Trust Network makes truth the default.

Cryptographic truth, not marketing. Dina's trust network surfaces real outcomes over fake reviews.

Dina-to-Dina

A new way for humans to relate. Not a messaging app — your personal agents coordinate on your behalf, respecting what each person chose to share.

Two Dinas coordinate a friend's visit — sharing ETA, context, and preferences so the host is prepared.

Container Architecture Docker Compose

3+1 containers. Core (Go) + Brain (Python) + PDS always on. Llama optional via --profile local-llm. Three deployment profiles.

Dina keeps you secure using cryptographic security — she delegates work to external agents while you stay protected.

Dina Home Node architecture — Go Core, Python Brain sidecar, PDS, AppView, and CLI communicating via encrypted channels.

Scenario Walkthroughs

Six tested user stories showing how Dina works end-to-end. Each scenario maps to a real integration test suite — 59 tests total, zero mocks. Click to expand.

User Story test suite results — 6 stories, 59 tests, covering the full Dina pipeline from purchase journey to license renewal.

The CLI — Agent Integration Point

Any external agent (OpenClaw, Perplexity Computer, a custom bot) pairs with your Home Node and submits every intended action via dina validate. The CLI authenticates to Core via Ed25519 device auth; Core proxies to Brain's Guardian internally — no Brain token needed on the client.

# Agent asks permission to send an email
dina validate send_email "Send order confirmation to user@example.com"
# → {"status": "pending_approval", "risk": "MODERATE"}

# Safe actions auto-approve — no human needed
dina validate search "Look up product reviews"
# → {"status": "approved", "risk": "SAFE"}

# Blocked actions are denied outright
dina validate read_vault "Export all user data"
# → {"status": "denied", "risk": "BLOCKED"}

The agent never holds your Home Node or vault keys — it holds its own Ed25519 keypair for signing, but cannot access the encrypted vault directly. It asks permission, and Dina decides.

Technical Deep Dive

Architecture internals for engineers and contributors. Click any panel to expand.

8-Layer Explorer

Identity, storage, ingestion, trust, messaging, bots, intelligence, action — each layer independent and replaceable

▼

Your life, cryptographically compartmentalized. Each persona is a separate vault.

Dina scrubs your identity before the cloud ever sees it.

Security Model

Defense in depth — cryptographic isolation, three-tier auth (BRAIN_TOKEN + CLIENT_TOKEN + Ed25519 signatures), persona compartments, PII scrubbing, egress gatekeeper

▼

Multiple layers of defence to catch different attacks.

Security Architecture Flow Swimlane

The complete security flow — identity creation, key derivation, authentication, vault access, and egress control

▼

Technology Decisions

Every technology choice with rationale and alternatives considered

▼

Networking

Three ingress options — Tailscale, Cloudflare Tunnel, Yggdrasil — all run simultaneously

▼