Hermes Agent — Guardian Token Protection Stack

A runaway AI agent can burn $300 in API credits overnight. A rate-limited agent goes silent for an hour and nobody notices. A long session runs out of context and starts hallucinating. Guardian is the proxy that prevents all three. This bundle deploys Hermes Agent on your own VPS, behind a Guardian proxy that handles everything production AI agents run into. Redis-backed token budgets so every session has a hard ceiling. Automatic model fallback when you hit rate limits — claude-haiku takes over, your agent keeps running. Conversation memory compaction before sessions hit context limits. A 30-minute heartbeat cron that alerts you the moment your agent goes silent. And a server hardened from day one with nginx TLS, Let's Encrypt SSL, UFW firewall, fail2ban, and daily backups. One install script. ~12 minutes. --- **What Guardian actually does:** - **Token budgets** — Set a daily and hourly limit per agent in `.env`. When it's hit, the agent returns a clear error instead of billing you. Hard ceiling, not a soft warning. - **Automatic model fallback** — When you hit a rate limit on Claude Sonnet, Guardian switches to claude-haiku automatically. If you have an OpenAI key, it can fall back further to gpt-4o-mini. No dropped requests, no manual intervention. - **Memory compaction** — Long conversations are summarized and compacted before they hit the context window limit. Your agent stays coherent across long sessions without losing track of what you discussed an hour ago. - **Heartbeat monitoring** — A cron job pings your agent every 30 minutes. No response? Telegram alert immediately. You know before your users do. - **Spend tracking** — Every API call logged. Daily spend reports via Telegram. Weekly summary. You always know what you're paying. --- **For non-technical buyers:** Think of Guardian as the manager your AI agent never had. It watches the budget, handles emergencies, keeps the agent running when rate limits hit, and reports back to you automatically. Without it, you're running an agent on autopilot with no guardrails — and no visibility into what it's spending. --- **Server hardening:** nginx with TLS termination, Let's Encrypt SSL, UFW firewall, fail2ban, SSH key enforcement, daily backups. --- **What you need:** - Ubuntu 22.04 or 24.04 VPS - Anthropic API key - Telegram bot token (optional — for alerts and spend reports) - Domain pointed at the server **Time to live:** 12 minutes.