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

A developer deployed two AI agents on a $7/month VPS using IRC as the transport layer. The public agent (nullclaw) is a lightweight 678 KB Zig binary using minimal RAM, accessible via web and IRC clients, while a private agent (ironclaw) handles email and scheduling over Tailscale. The system uses tiered inference with Claude Haiku for conversation and Sonnet for tool use, maintaining a hard $2/day cost cap through A2A passthrough architecture that consolidates API keys and billing.

Key Points

• •Deploy lightweight AI agents in Zig (~678 KB binary, ~1 MB RAM) for extreme cost efficiency on minimal infrastructure
• •Use IRC as a universal transport layer—enables both web clients (gamja) and native IRC client access simultaneously
• •Implement tiered inference strategy: fast/cheap models (Haiku 4.5) for conversation, capable models (Sonnet 4.6) only for tool use
• •Enforce hard cost caps ($2/day) by rate-limiting and model selection to prevent runaway API bills
• •Separate public and private agents across boxes with Tailscale + Google A2A protocol for secure, authenticated access
• •Consolidate billing through A2A passthrough: private agent reuses gateway's inference pipeline, single API key for all requests
• •Embed IRC client (gamja) directly in website for frictionless user access without requiring external IRC client setup
• •Use TLS on IRC (port 6697) for encrypted communication between clients and server

Workflow Diagram

Concepts