Hermes Agent — Architecture

Self-hosted AI operations assistant · architecture overview

SELF-MANAGED VPS (Docker, firewall: SSH only) TelegramBot API, outboundAllow-listed usersApprove in chat:Allow once / DenyHermes AgentDocker container, auto-restartProfiles: default, n8n-engineerTools and skills (files, shell)Memory kept on the VPSCommands that need permissionwait for my approvalNo SSH and no Docker socketaccess grantedn8n (demo instance)Separate from the Kosan botSample workflows onlyReached through MCPAgent drafts workflowsI review the JSON before useNo production dataModel gateway (9router)Local, OpenAI-compatibleFallback order across modelsUsage log per requestAPI keys in env files,never shared with the agentHermes dashboardAdmin UI for config,channels, MCP and profilesTailnet only, behind loginIsolationOwn folder and compose fileCannot see the Kosan bot,its Sheet or its secretsInternal Docker networking;nothing published to internetModel providersFree-tier routesMy own API keyChosen per task throughthe gateway; the agent isunaware of which providerchat APIMCP

Control

  • Tasks are delegated from a phone through Telegram.
  • Risky commands pause for Allow once or Deny.
  • No server access: the agent works inside its own container.

Routing

  • One local gateway in front of all providers.
  • Fallback order, so a rate limit does not stop the agent.
  • Per-request usage is visible, which matters when credit is paid.

n8n via MCP

  • The agent drafts workflows on a demo instance only.
  • I read the generated JSON, not the agent's summary.
  • Production workflows stay out of its reach.