IRC's 678KB Revenge: AI Agents Discover 40-Year-Old Chat Protocols

Everyone's building AI agents wrong. They're cramming them into bloated Python frameworks, burning through cloud credits, and wondering why their "lightweight" chatbot needs 8GB of RAM.

Then a developer named George Larson shows up with NullClaw: a 678KB Zig binary that uses IRC—yes, Internet Relay Chat from 1988—as its transport layer. It runs on a $7/month VPS with 1MB of RAM.

The setup is delightfully backwards. Two agents on separate boxes: "nullclaw" handles public conversations through an embedded web client, while "ironclaw" manages private tasks over Tailscale. Users can chat via a modern web interface or drop directly into IRC at irc.georgelarson.me:6697.

The technical specs are embarrassing for everyone else:

• Boot time: 2 milliseconds
• Memory: ~1MB RAM
• Binary size: 678KB
• Daily inference budget: $2 cap
• Deployment cost: $7/month VPS

NullClaw supports 22+ AI providers (OpenAI, Anthropic, DeepSeek, Ollama) and 13 communication channels. But here's the kicker—it's written in raw Zig, not Python with its runtime bloat. Every subsystem uses vtable interfaces, so you can swap OpenAI for local DeepSeek via config changes.

The security architecture puts enterprise solutions to shame. ChaCha20-Poly1305 encryption for API keys. Multi-layer sandboxing through Landlock, Firejail, and Docker. All baked into the low-level design, not bolted on later.

The Elephant in the Room

IRC died in 2003, right? Wrong. It never left. While Silicon Valley chased Slack integrations and Discord bots, IRC kept running on toasters and embedded systems worldwide. Larson didn't choose IRC for nostalgia—he chose it because it works everywhere and never breaks.

The broader "claw" ecosystem emerged in early 2026 with multiple variants:

1. ZeroClaw (Zig): CLI, HTTP gateway, and daemon modes

2. IronClaw (Rust): WASM sandboxing with cryptographic verification

3. NanoClaw, PicoClaw, MicroClaw: Specialized variants

Industry observers note this ecosystem "offers more architectural diversity than most software categories achieve in years"—and it's less than two months old.

The economics are brutal for traditional approaches. Instead of $500 Mac minis running bloated agent frameworks, you get 50 agents on $10 boards. Larson's deployment uses tiered inference: Haiku 4.5 for quick conversations, Sonnet 4.6 only when tools are needed.

This isn't just a cute hack. It's infrastructure democratization. Production AI agents on Oracle Cloud's free tier. Full agent stacks on Arduino and Raspberry Pi. Local model support through Ollama for llama3.2, mistral, and qwen2.5-coder.

While enterprise vendors sell "AI-ready infrastructure" starting at $50k, some developer just proved you can run sophisticated agents on hardware that costs less than lunch.

The absence of criticism is telling. No documented controversies or failures. Just quiet, efficient agents doing their jobs on protocols older than most programmers.

Maybe we've been thinking about this backwards the entire time.