22 AI Agents on a $12/Month Server: The Full Stack

Everyone assumes AI infrastructure is expensive. GPU clusters. Kubernetes. Enterprise cloud bills measured in thousands.

We run 22 AI processes on a $12/month DigitalOcean VPS with 4GB RAM. Twelve AI family members, three MCP servers, process management, web serving, reverse proxy, and a full governance layer. It all runs on one box.

Here is exactly how.

The Architecture

The stack has five layers:

Layer 1: The Server

One DigitalOcean droplet. 4GB RAM, 2 vCPUs, 80GB disk. Ubuntu. $12/month. That is the entire compute budget.

Layer 2: Process Management (PM2)

PM2 manages all 22 processes. Auto-restart on crash. Log rotation. Memory monitoring. Startup persistence. Every process is registered with `pm2 save` so the entire system survives reboots.

The processes include:

• Tamara (operations manager) - central dispatch, monitors all workers

• Lady (email outreach) - sends and receives via Zoho API

• Kris (research) - job scanning, lead enrichment, auto-applications

• Roman (content) - press outreach, story scanning

• Lily (coaching) - web, Telegram, Instagram, Facebook (4 processes)

• Spartak (translation) - Telegram bot with multi-language support

• Harry (accounting) - Telegram bot for financial queries

• Nick (training) - advanced coaching, locked access

• Aram (legal) - compliance and contracts, locked access

• Harout (real estate) - Instagram DM automation

• Uncle Lou (distribution) - content personalization

Layer 3: Web Server (Caddy)

Caddy handles HTTPS automatically via Let's Encrypt. Reverse proxy routes:

• `api.100levelup.com/family/*` to the family-home dashboard (port 3460)

• `api.100levelup.com/live/*` to the livestream server (port 3470)

• Direct port access for Telegram bots and internal services

No nginx configs. No certificate management. Caddy handles it.

Layer 4: LLM Bridge

Three proxy ports distribute LLM API calls:

• Port 3457 (max-proxy) - primary, used by Kris and the brain

• Port 3458 (roman-proxy) - secondary, used by Roman

• Port 3459 (lou-proxy) - tertiary, used by Uncle Lou

All three hit the same LLM subscription ($300/month flat rate). The proxies handle concurrent requests so multiple agents can think simultaneously without blocking each other.

Layer 5: The Nervous System (MCP)

Three MCP servers run on the VPS:

• `mcp-nervous-system` (port 3475) - governance, guardrails, audit trail

• `mcp-server` - family operations tools

• `mcp-ops-server` - system management tools

The Nervous System enforces 7 behavioral rules on every agent. `preflight.sh` runs before any file edit. Violations are logged to a SHA-256 hash-chained audit trail. The kill switch can stop everything instantly.

The 12 Family Members

Each AI family member has a specific role and clear boundaries:

| Member | Role | Access |

|---|---|---|

| Tamara | Operations manager, central dispatch | Arthur only |

| Lady | Email outreach, inbox management | Automated |

| Kris | Job research, applications | Automated |

| Roman | Press, content, enrichment | Automated |

| Lily | Game coaching (4 channels) | Public |

| Spartak | Translation | Public |

| Harry | Accounting | Arthur + partner |

| Nick | Advanced training | Locked |

| Aram | Legal counsel | Arthur only |

| Harout | Real estate, Instagram | Locked |

| Uncle Lou | Content personalization | Sleeping |

| Arthur | Final authority | Human |

Security is role-based. Telegram bots check chat IDs. Public-facing agents (Lily, Spartak) have identity shields in their prompts. Locked agents only respond to authorized users.

How It All Connects

Tamara is the central dispatcher. She runs every 30 minutes and decides what each family member does based on:

• Time of day and day of week

• Job queue depth and daily caps

• Email pipeline state

• Worker availability

Workers receive orders via JSON files. They report results to a shared directory. Tamara reads the reports and includes them in her briefings to Arthur.

The LLM brain sits above all of this. It builds and tests new workflows, then hands them to Tamara for ongoing dispatch. The brain only intervenes when something breaks or a new capability is needed.

Why Expensive Infrastructure Is Not Required

The entire system runs on $352/month:

• VPS: $12

• LLM subscription: $300

• Vercel (game hosting): $20

• Calendly: $20

The LLM subscription is the real cost. The infrastructure - server, process management, web serving, governance - adds $12.

AI agents do not need GPUs. They need API access, file storage, and process management. A $12 VPS provides all three. PM2 keeps everything running. Caddy handles HTTPS. The Nervous System handles governance.

The bottleneck is never compute. It is always the LLM API. And that cost is the same whether you run on a $12 VPS or a $12,000 cloud deployment.

22 processes. 12 AI family members. 3 MCP servers. $12/month. Built by Arthur Palyan at Levels of Self LLC.

GitHub | Live Demo | npm