Show HN: I put an AI agent on a $7/month VPS with IRC as its transport layer

This post describes building a portfolio AI agent system using IRC as the transport layer, deployed on a $7/month VPS with a two-agent architecture and strict security boundaries.

• •Two agents run on separate boxes: nullclaw (public-facing, 1MB RAM Zig binary) handles visitor questions and clones repos to verify claims, while ironclaw (private box via Tailscale) handles email, calendar, and sensitive context
• •Tiered model selection keeps costs under $2/day: Claude Haiku 4.5 for conversational triage, Sonnet 4.6 only when tool-use or deep reasoning is needed
• •IRC (ergo server + gamja web client) was chosen for full stack ownership, zero vendor lock-in, and aesthetic fit with a terminal-style portfolio
• •Google's A2A protocol (v0.3.0) enables agent-to-agent calls over Tailscale; ironclaw's nullclaw instance proxies inference through ironclaw's own gateway so only one API key and billing relationship is needed