Architecture

5-layer Anthropic-native engine architecture for Databayt

Architecture

Kun is a configuration engine — not a server. It sits as the configuration layer on top of Anthropic's product suite, running both technical and business sides of Databayt.

5-Layer Architecture

Layer 5: Company Operations
  Cowork | Claude Apps | Enterprise Connectors
  Ali (business) | Samia (content) | Team coordination

Layer 4: Coordination & Automation
  Agent Teams | Scheduled Tasks | CI/CD | Channels

Layer 3: KUN CONFIGURATION ENGINE (core value)
  CLAUDE.md | 40 Agents | 25 Skills | 18 MCP Servers
  8 Rules  | 5 Hooks   | 6 Memory  | 100+ Keywords

Layer 2: Developer Surfaces (Anthropic-provided)
  CLI | VS Code | JetBrains | Desktop | Web | iOS

Layer 1: Foundation (Anthropic-provided)
  Opus 4.6 | Sonnet 4.6 | Haiku 4.5 | 1M Context | API

Team Roles

Member	Role
Abdout	Builder
Ali	QA Engineer + Sales
Samia	R&D & Kun Caretaker
Sedon	Executor

Layer 3: The Engine (Kun's Core)

CLAUDE.md Hierarchy

Priority (High → Low):
1. Project-level  ~/project/CLAUDE.md         — Project context
2. Repo-level     ~/project/.claude/CLAUDE.md  — Keywords, workflows
3. User-level     ~/.claude/CLAUDE.md          — Stack, preferences
4. Pattern library /Users/abdout/codebase/     — Core patterns

Agent Fleet (40 agents, 4 tiers)

captain (Tier 0 — company brain)
├── Business (Tier 1): revenue, growth, support
├── Product (Tier 2): product, analyst
└── Tech Leadership (Tier 3): tech-lead, ops, guardian
        └── orchestration → 31 specialist agents

Tier	Agents	Domain
Tier 0 (1)	captain	CEO brain, weekly allocation, revenue strategy
Tier 1 (3)	revenue, growth, support	Business operations
Tier 2 (2)	product, analyst	Product roadmap, market intelligence
Tier 3 (3)	tech-lead, ops, guardian	Architecture, infra, security
Stack (7)	nextjs, react, typescript, tailwind, prisma, shadcn, authjs	Technology expertise
Design (4)	orchestration, architecture, pattern, structure	System design
UI (4)	shadcn, atom, template, block	Component hierarchy
DevOps (3)	build, deploy, test	Development lifecycle
VCS (2)	git, github	Version control
Specialized (8)	middleware, i18n, semantic, sse, optimize, performance, comment, icon	Domain expertise
Reference (4)	hogwarts, souq, mkan, shifa	Product-specific patterns

MCP Ecosystem (18 servers)

Category	Servers
UI & Design	shadcn, figma, tailwind, a11y, storybook
Testing	browser (headless), browser-headed
DevOps	github, vercel, sentry, gcloud
Data & Auth	neon, postgres, stripe, keychain
Knowledge	ref, context7
Project Mgmt	linear

Rules Engine (8 rules)

Auto-activate on file patterns: auth, i18n, prisma, tailwind, testing, deployment, multi-repo, org-refs.

Hook Automation (5 hooks)

Event	Action
SessionStart	Print model + timestamp
PreToolUse (pnpm dev)	Kill port 3000
PostToolUse (pnpm dev)	Open Chrome
PostToolUse (Write/Edit)	Run Prettier
Stop	Log session end

Layer 4: Coordination

Agent Teams (Experimental)

Lead agent coordinates teammate agents, each in isolated git worktrees. No merge conflicts. Consolidated PR.

Cross-Device Flow

Phone (iOS) ──Dispatch──▶ Desktop App ──/teleport──▶ Terminal (CLI)
     ◀──────── Remote Control ──────────────────────────────┘

Scheduled Tasks

Type	Runs On	Use Case
Cloud	Anthropic infra (computer off)	Health checks, dependency updates
Desktop	Local (app open)	Recurring builds
/loop	Active session	Poll deploy status

3-Channel Communication

Channel	Medium	Direction	Purpose
Dispatch	Apple Notes → Dispatch (Captain, Cowork, Inbox)	Async, bidirectional	Updates, decisions, handoffs
GitHub Issues	databayt/*/issues	Structured work items	Tasks with labels, milestones, assignees
Claude Native	Code / Cowork / Voice	Real-time	Building, planning, quick decisions

Cowork ↔ Code Bridge

Cowork (think): plan → create issues → write to Dispatch/Cowork
Code (do):      read Dispatch/Cowork → pick up issues → execute → report back

Shared state: ~/.claude/ (agents, memory, settings, rules). Same brain, two modes.

Layer 5: Business Operations

Team Workflows

Member	Role	Primary Workflows
Abdout	Builder	dev → build → push → deploy
Ali	QA + Sales	Testing, issue reports, outreach (sales@databayt.org)
Samia	R&D	Claude/Anthropic research, sharing economy, Kun care
Sedon	Executor	Clear task maps, Saudi operations

Hogwarts Pilot (King Fahad Schools)

Developer: handover admission → 5-pass QA → deploy to ed.databayt.org
Business:  Cowork → onboarding plan → Arabic welcome email → training agenda
Content:   Cowork → Arabic user guide → screenshots → voiceover

Repository Map

databayt/ (14 repos)
├── Products
│   ├── hogwarts     — Education SaaS (FLAGSHIP, daily commits)
│   ├── mkan         — Rental marketplace (Phase 1 done)
│   ├── souq         — E-commerce (MVP, paused)
│   └── shifa        — Medical platform (early, paused)
├── Libraries
│   ├── codebase     — Pattern library (54 ui + 62 atoms + 31 templates)
│   ├── shadcn       — UI components (shadcn/ui fork)
│   └── radix        — UI primitives (Radix fork)
├── Engine
│   └── kun          — Configuration engine (this repo)
├── Mobile & Marketing
│   ├── swift-app    — iOS app (Swift 6/SwiftUI)
│   └── marketing    — Landing pages
└── Other
    ├── spma         — Internal tool
    ├── apple        — Design R&D
    ├── distributed-computer — Rust/blockchain concept
    └── .github      — Org profile

Security

Layer	Mechanism
Anthropic	SSO/SCIM, audit logging, SOC 2, no training on data
Kun config	4 deny rules (rm -rf, DROP TABLE, etc.), 38 explicit allow rules
Hooks	PreToolUse guards validate before execution
Secrets	macOS Keychain MCP, never in git
Accessibility	Screen reader compatibility, VoiceOver testing

Architecture Decision Records

ADR-001: Configuration Over Infrastructure

Decision: Use Anthropic's products as-is, configure don't build. Rationale: 4-person team can't maintain custom infrastructure alongside 5 products.

ADR-002: Opus 4.6 Default

Decision: Use Opus for all agents, Haiku only for explore subagents. Rationale: Quality over speed for a small team where every output matters.

ADR-003: Single Subscription Model

Decision: One Max 20x ($200/month) shared via Desktop/Web. Rationale: $500/month burn with $5K capital = 10 months runway. Add Pro seats ($20) when revenue grows.

ADR-004: Hogwarts-First Development

Decision: All engine features validated through Hogwarts pilot first. Rationale: King Fahad Schools is the revenue pipeline. Ship what pays.

ADR-005: Shared Component Library

Decision: All products pull from databayt/codebase. Rationale: 54 UI + 62 atoms + 31 templates = don't rebuild across 5 products.

Architecture Principles

Configuration over infrastructure — Don't build what Anthropic ships
Architecture-first — Humans design, AI generates within constraints
Guardrails as training data — CLAUDE.md, agents, rules shape output
Full spectrum — Technical AND business operations
Anthropic-native — Use products as designed

Configuration Workflows