DAO
Back

automation

Hermes Agent — Self-Hosted AI Operations Assistant

A personal AI agent running in Docker on my own VPS. I delegate small tasks from Telegram, approve any command that needs permission, and let it draft n8n workflows on a demo instance that I review before anything is used.

Role
Solo: deployment, container setup, integrations, access boundaries
Period
Nov 2026 – present
Stack
Hermes AgentDockerTelegram Bot APIModel routingMCPn8n

Results

  • Runs in Docker on my own VPS and restarts on its own after a reboot
  • Tasks are delegated from Telegram; commands that need permission wait for my Allow once or Deny
  • Model calls go through a local gateway with a fallback order, so switching providers does not change the agent setup
  • Connected to a demo n8n instance through MCP to draft workflows; I review the JSON before it is used
  • Keys stay in environment files on my server, and the agent has no access to the Kosan bot or its data

Architecture

Problem

Running automation means a long tail of small jobs: drafting a workflow, checking a config, writing an update. An AI agent can take part of that load, but only if it is useful without being trusted blindly.

What I built

  • Telegram gateway so tasks can be delegated from a phone.
  • Docker deployment with a restart policy, so the agent survives reboots.
  • Model routing through a local OpenAI-compatible gateway with a fallback order and per-request usage.
  • MCP link to a demo n8n instance and a separate agent profile for workflow drafting.
  • Access boundaries: its own folder and compose file, an admin dashboard reachable only on my private network, and no SSH or Docker socket access.

Challenges and decisions

  • Review the output, not the report. A workflow drafted by a free model looked complete but had many bugs: a key written into a node, an invented business rule, no protection against sending twice. I now read the generated JSON instead of trusting the agent’s summary.
  • Say no to access. The agent asked for server access and raw API keys. I declined both, and gave it only the tools its job needs.
  • Watch the cost. Every tool call resends the whole conversation, so long sessions get expensive fast. I track usage per request and keep sessions short.