Agentic hub: from OpenClaw to Hermes
My own always-on AI agents, on a private server and reached through Telegram. It started on OpenClaw in February 2026 and moved to Hermes in June 2026. An appointment agent books and cancels real appointments with an email inbox of its own, and a fitness agent coaches me after every gym session I log.
- When
- Feb 2026 to now
- Status
- Online Running, for my own use.
- Where
- Own setup, running around the clock on my own server
- Built with
- Hermes Agent, OpenClaw, OpenRouter, Telegram, AgentMail, Browser automation, Gmail
- Made as
- Individual project
About the project
I run my own AI agents on a private server that is reachable only through my private network. I talk to them through Telegram, where each topic (appointments, gym, business, development) has its own conversation, and some have an agent of their own. The models run through OpenRouter, so I can switch between them.
How it grew
- 7 February 2026: set up the server and OpenClaw, my first agent platform.
- 17 and 18 March 2026: an email inbox of the agent's own (AgentMail), browser automation, Google Calendar and Gmail.
- 21 March 2026: the first appointment agent, for a barbershop.
- 7 April 2026: a planner that finds back-to-back slots at two places and books both.
- 28 April 2026: a second appointment agent, for a salon booked through Treatwell.
- 9 June 2026: moved everything to Hermes. Its migration tool carried over 21 items (the Telegram channel, memory, identity and skills) with no errors.
- June 2026: my own personalised workout logger for the gym; I shared each session with the fitness chat in Telegram.
- 13 August 2026: the logger was connected to the fitness agent, which now answers every session in Telegram on its own.
What the agents do
- Appointment agent: checks availability, books, cancels and keeps track of appointments at a barbershop and a salon by working their booking websites in a browser. Confirmations and cancel links arrive in the agent's own AgentMail inbox, so cancelling takes one message to the agent. Asked for a "beauty day", it finds back-to-back slots at both places and books both.
- Email triage: sorts my inbox by priority and drafts replies in my voice; nothing is sent until I approve it. It works with Gmail today, and the triage itself does not depend on the provider, so other mailboxes can be added.
- Fitness: I built a personalised workout logger that I use at the gym. After each session, my fitness agent looks at how it went and sends me a short coaching message in Telegram.
- In progress An editorial team of three agents (a writer, an editor and a critic) for long-form articles, coordinated on a Kanban board, with my approval at every stage.
How it is kept safe
- The server, and the workout logger on it, are reachable only through my private network, not from the open internet; the logger also needs a token.
- Commands that change something need my approval, and scheduled jobs cannot run them at all.
- API keys and passwords are masked before a command's output reaches the agent or its logs, so an agent cannot repeat them in a reply.