name: AG Dev Hub Architecture Plan description: Full architecture plan for multi-agent infrastructure — Claude Code + Gemini CLI on VPS, Telegram control, LLM Wiki shared memory, n8n workflows, voice assistant future type: project

Architecture Overview

┌──────────────────────────────────────────────────┐
│                   VPS (HOSTiQ kVPS70)             │
│                                                    │
│  ┌──────────────┐        ┌──────────────┐         │
│  │ Claude Code   │──────▶│ Gemini CLI    │         │
│  │ (orchestrator)│  delegates│ (heavy dev) │         │
│  │ Max/Sonnet4.6 │◀──────│ Ultra/Gemini3│         │
│  └──────┬───────┘        └──────────────┘         │
│         │                                          │
│  ┌──────┴───────┐   ┌──────────────┐              │
│  │ Channels      │   │ LLM Wiki     │              │
│  │ (Telegram)    │   │ (git repo)   │              │
│  └──────┬───────┘   └──────┬───────┘              │
│         │                   │                      │
│  ┌──────┴───────┐   ┌──────┴───────┐              │
│  │ n8n          │   │ Caddy        │              │
│  │ (workflows)  │   │ (reverse     │              │
│  │              │   │  proxy+SSL)  │              │
│  └──────────────┘   └──────────────┘              │
└────────────────────────┼──────────────────────────┘
                          │ git sync
          ┌───────────────┼───────────────┐
          │               │               │
          ▼               ▼               ▼
    ┌──────────┐  ┌──────────┐    ┌──────────┐
    │ Cowork    │  │ Obsidian  │    │ Telegram  │
    │ (desktop) │  │ (mobile+  │    │ (phone)   │
    │           │  │  desktop) │    │           │
    └──────────┘  └──────────┘    └──────────┘

Role Distribution

• Claude Code (Max, Sonnet 4.6): orchestrator, Telegram interface, wiki maintenance, planning
• Gemini CLI (Ultra, Gemini 3): heavy coding tasks, large context analysis, delegated by Claude Code
• n8n: scheduled workflows, API connectors, data pipelines (Pediatric News, Ad Analytics, Zest Content)
• Caddy: HTTPS reverse proxy with auto SSL for web services

Google Accounts Mapping on Server

• Personal (gmail) → Claude Code OAuth
• Subscription → Gemini CLI OAuth (AI Ultra for Business)
• Work (corporate) → gcloud for GCP resources ($100/mo budget)

Shared Memory Strategy (LLM Wiki pattern by Karpathy)

• Markdown wiki in /srv/wiki/ with cross-links
• Git repo synced to GitHub → pulled by Cowork and Obsidian
• Each project has its own directory: lunar-hubble, ad-analytics, pediatric-news, zest-content
• Claude Code auto-updates wiki pages after significant work
• Future: add LightRAG for semantic search when wiki exceeds 100 pages

Budget

• Claude Max: $100/mo
• Google AI Ultra: already subscribed
• HOSTiQ kVPS70: 815 грн/mo (~$19)
• Total: ~$119/mo

Future Additions

• Voice layer: faster-whisper (STT) + ukrainian-tts (TTS) for Telegram voice messages
• Requires VPS upgrade to kVPS90 (8 GB RAM) for ~$25/mo
• Obsidian as wiki viewer on desktop and mobile
• LightRAG MCP for semantic search over wiki