Architecture

System design, components, and how everything fits together.

System Overview

┌─────────────────────────────────────────────────────────────┐
│                        Nginx (SSL)                          │
│  k8.virtualgpt.cloud → static    meet.virtualgpt.cloud →   │
│                                  static + /api/ + /ws       │
└──────────────┬─────────────────────────┬────────────────────┘
               │                         │
┌──────────────▼──────────┐  ┌───────────▼──────────────────┐
│   K8s Landing Page      │  │      GMeet Frontend          │
│   (Static HTML)         │  │  (React + Tailwind + Zustand)│
└─────────────────────────┘  └──────────┬──────────┬────────┘
                                        │ REST     │ WS
                             ┌──────────▼──────────▼────────┐
                             │      GMeet Backend            │
                             │  Express :4000 + WS :4001     │
                             └──┬─────┬─────┬─────┬─────┬───┘
                                │     │     │     │     │
                             ┌──▼─┐┌──▼─┐┌──▼─┐┌──▼─┐┌──▼─┐
                             │5001││5002││5003││5004││5005│
                             │BMAD││Mary││John││Fred││Dana│
                             └────┘└────┘└────┘└────┘└────┘
                              Agent Zero Containers / Pods
                                        │
                                   OpenRouter API
                              (Kimi K2.5 / Gemini Flash)

Components

Agent Zero Instances

Each agent runs in an isolated Docker container (or K8s pod) using the agent0ai/agent-zero image. Each instance has:

• Its own agent profile (persona, system prompt, skills)
• Tuned model parameters (temperature, context length)
• Shared project volume for collaboration artifacts
• A2A server enabled for inter-agent communication
• Web UI on port 80 (mapped to 50001–50009 externally)

GMeet Backend

An Express.js server that bridges the GMeet frontend with Agent Zero instances:

• REST API on port 4000 — meeting management, agent listing, message routing
• WebSocket server on port 4001 — real-time transcript streaming
• X-API-KEY auth — authenticates with each Agent Zero instance using the mcp_server_token
• Context management — maintains conversation context IDs for multi-turn dialogues

GMeet Frontend

A React single-page application built with Vite:

• Lobby — select agents to invite to a meeting
• AgentGrid — Google Meet-style grid showing active agents with avatars
• TranscriptPanel — real-time conversation transcript with agent color coding
• Controls — message input, send to all or specific agents
• State management via Zustand with WebSocket sync

Nginx Reverse Proxy

Nginx handles SSL termination and routing:

• k8.virtualgpt.cloud → static files from /var/www/k8.virtualgpt.cloud/
• meet.virtualgpt.cloud → static frontend + proxy /api/ to :4000 + proxy /ws to :4001
• Let's Encrypt SSL certificates with auto-renewal

Data Flow

1. User opens meet.virtualgpt.cloud → loads React SPA
2. Frontend fetches agent list via GET /api/agents
3. User creates a meeting → POST /api/meetings
4. User sends a message → POST /api/meetings/:id/message
5. Backend forwards message to selected agents via POST /api_message with X-API-KEY
6. Agent Zero processes message using configured LLM (via OpenRouter)
7. Response is returned to backend → stored in transcript → broadcast via WebSocket
8. Frontend updates transcript panel in real time

Model Configuration

Purpose	Provider	Model
Chat (reasoning)	OpenRouter	moonshotai/kimi-k2.5
Utility (fast tasks)	OpenRouter	google/gemini-2.5-flash-preview
Embeddings	OpenAI	text-embedding-3-small

Models are configured via the a0-dotenv/.env file mounted into each container. See Configuration for details on switching models.

Repository Structure

agent0-bmad-k8/
├── agents/                    # Agent profiles (9 agents)
│   ├── bmad-orchestrator/     # settings.json, prompts/, SKILL.md
│   ├── bmad-analyst/
│   ├── bmad-pm/
│   ├── bmad-architect/
│   ├── bmad-ux/
│   ├── bmad-po/
│   ├── bmad-sm/
│   ├── bmad-dev/
│   └── bmad-qa/
├── gmeet/
│   ├── backend/               # Express API + WebSocket server
│   └── frontend/              # React + Tailwind + Zustand SPA
├── k8s/bmad/                  # Kubernetes manifests
│   ├── namespace.yaml
│   ├── configmap.yaml
│   ├── secrets.yaml
│   ├── deployment-template.yaml
│   ├── pvc.yaml
│   ├── ingress.yaml
│   └── deploy-bmad.sh
├── projects/bmad-template/    # BMAD workflow templates
├── docker-compose.live.yml    # Live demo (5 agents)
├── docker-compose.bmad.yml    # Full setup (9 agents)
├── a0-dotenv/.env             # API keys (gitignored)
├── .github/workflows/ci.yml   # GitHub Actions CI
└── README.md

Security

• API keys are never committed to git — stored in .env.live and a0-dotenv/.env (both gitignored)
• Agent auth uses X-API-KEY header with per-instance tokens
• Web UI auth uses session-based login (configurable credentials)
• K8s secrets store API keys as base64-encoded Kubernetes secrets
• SSL/TLS via Let's Encrypt with auto-renewal