The problem
Hosted bot platforms (Intercom Fin, Zendesk Answer Bot) charge per resolution. Self-hosting replaces that fee with a server and an LLM bill, and keeps customer data off third-party AI infrastructure.
Recommended setup
| Agent | Hermes Agent + Telegram/Discord/web channel via custom MCP tools |
| Hardware | Hetzner CCX13 (€38/month, dedicated CPU, 16 GB RAM), shared CX22 isn't enough for sustained traffic |
| LLM | Claude 4.5 Sonnet for quality + OpenRouter fallback when rate-limited |
How it works
Hermes Agent indexes your help docs into a RAG store on startup. Customer-facing chat widget (or Telegram bot) sends queries to the agent. Agent retrieves relevant doc chunks, generates answer with citations, escalates to human via webhook if confidence below threshold or specific keywords detected (refund, complaint, legal).
Reality check
We have not run this on a live support queue, so there is no resolution rate or satisfaction score to quote. Measure two things before you rely on it: the share of the tickets it closes that stay closed, and how fast an escalation reaches a human. Keep refunds, complaints and anything account-specific with people.
What breaks
- Account-specific questions (the agent shouldn't see private customer data: escalate)
- Complaints (always escalate: people handle these better and customers expect it)
- Documentation gaps (agent will fail confidently on questions not covered)
Alternative setups
ClawRift ($19 or $49 a month) or NitroClaw ($20 or $100 a month) if you don't want the ops work, at the prices on their own pricing pages on 18 October 2026. Trade-off: data goes through their infrastructure.