I Put an AI Agent on a $7 VPS (IRC Is the Secret Weapon)

Every portfolio site with an AI chatbot does the same dumb thing: feed a resume into an LLM and let visitors rephrase it. It's a parlor trick. The model can't tell you anything the resume doesn't already say.

George Larson wanted something different. When a hiring manager asks "how does George handle test coverage?" the answer shouldn't be "George values comprehensive testing." It should clone the repo, count the tests, read the CI config, and come back with specifics.

So he built it. On a $7/month VPS. Using IRC as the transport layer.

The Architecture

The setup is beautifully simple — two agents, two boxes, two security boundaries:

visitor (browser)
    → gamja web IRC client
    → wss://nullclaw.georgelarson.me:443
    → Cloudflare (proxy + TLS)
    → ergo IRC server
    → #lobby channel
    → nullclaw agent (public-facing, reads GitHub repos)
    → #backoffice (private channel)
    → ironclaw agent (private box, has email/calendar)

The public agent (nullclaw) runs on a minimal perimeter box. A 678KB Zig binary using about 1MB of RAM. It handles greetings, answers questions about projects, and clones repos to substantiate claims with real code.

The private agent (ironclaw) sits on a separate system via Tailscale. Different box. Different security boundary. The public box has zero access to private data.

Why IRC?

He could have used Discord, Telegram, or built a custom WebSocket chat. But IRC wins for three reasons:

1. It fits. Terminal-style portfolio → IRC client embedded. Discord would feel wrong.

2. You own everything. Ergo server, gamja client, nullclaw agent — all on your infrastructure. No API that changes its terms, no platform that deprecates bot access.

3. It's 30 years old and bulletproof. IRC is simple, well-understood, zero vendor lock-in.

The same agent that talks to visitors via the web client can talk to you via irssi from a terminal. That's elegant.

Model Tiering: The Real Cost Hack

Here's where most people screw up. They reach for the biggest model they can afford. Wrong instinct for a doorman.

Using Opus for a concierge would signal the opposite of model understanding. If Haiku can handle it, don't send it to Sonnet. Tiered inference is how you keep this under $2/day.

Security (Because It's a Public Box)

This isn't a toy — it's a public-facing perimeter. Hardened like one:

• SSH: Key-only auth, non-standard port, root disabled
• Firewall: UFW with only 3 ports (SSH, IRC/TLS, HTTPS)
• Cloudflare proxy: Visitors never hit the box directly
• Agent sandboxing: Workspace-only file access, read-only tools, 10 actions/hour max
• Cost controls: $2/day, $30/month hard caps
• Audit logging: Every tool call logged

The philosophy is minimal blast radius. If it gets compromised, the damage is an IRC bot with a $2/day inference budget.

The Numbers

Total footprint:
  → Under 10 MB of binaries
  → Under 5 MB of RAM at idle
  → $7/month VPS
  → $2/day inference (capped)
  → ~$67/month all-in

For less than the cost of a Netflix subscription + a few coffees, you get a fully functional AI agent that can talk to visitors, read your code, and answer questions with real substance.

What This Means for Indie Hackers

You don't need a $500/month GPU server to run a useful AI agent. You don't need Discord's API or Telegram's bot platform. You don't need the latest frontier model for every conversation.

You need:

1. A cheap VPS ($7/month gets you plenty)
2. An old protocol that still works (IRC, baby)
3. Smart model tiering (cheap for the hot path, capable for heavy lifting)
4. Proper security boundaries (different boxes for public vs private)

The entire stack is small, self-hosted, and replaceable. No vendor lock-in. No API deprecation anxiety. Just you, your code, and a 678KB binary doing work while you sleep.

Original source: georgelarson.me — Building a Digital Doorman