AI Systems
Setup

Many owners have already bought AI: Claude, Copilot, an agent framework. They started several projects and finished none: a bot goes silent, memory does not stick, the automations were built by someone who left. I work in these tools every day and run a system on my own server that I set up for clients the same way

On my server 8 Claude Code agents run as plain services, and a clinic owner has 6 more. Before every answer an agent loads what it needs from a memory graph of 8,274 nodes, and after the answer it stores what is new. 16 timers and a night watcher look after everything: in 276 night runs it found a failure 74 times and sent a card with buttons. I have set up personal Hermes assistants for six more people

For a similar solution for your business, see Support & Growth

What You Get

If you already have AI tools that work only now and then, I start with an inventory: what runs, what is broken, what spends money. Then I bring everything to a working state, put guards around it and teach you to run the system from Telegram with words and buttons, no terminal

How It Works

Access to the server and accounts in read mode. Day one is an inventory: what runs, what is broken, where the keys are, what spends money. Nothing changes without a backup and a log entry

Every agent is Claude Code in channels mode with the Telegram plugin, a separate service with its own folder and secrets. Updates come from the vendor, not from hand-patching

If the plugin had to be changed, the change lives as a reference with a fingerprint. Every 30 minutes a guard compares the live file with the reference and restores it if they differ

The memory graph lives in a database: before an answer the agent loads what it needs, after the answer it stores what is new. My server has 7,279 such reads in its log

Watchdogs check the agents every 5 minutes, a shared reply guard re-sends the answer if a bot stays silent, the night watcher sends a card with buttons if something broke, and a separate guard checks the watcher itself. A spending report arrives every night

Every change on the server is a log entry: 829 of them since April 2026, plus 149 independent audit reports

Tools

• Claude Code

  the core of every agent, channels mode

• MCP

  connects agents to outside services

• Hermes

  a second engine for a personal assistant, with a fallback model

• Hooks

  memory before and after an answer, a guard for silent replies

• Linux server

  a service per agent, timers and guards

• Telegram

  control with words and buttons

Result

As of October 22, 2026, 8 of 8 agents run on my server and 6 of 6 for the clinic owner, and Hermes assistants, mine and two others, were checked live. I do not promise that nothing will ever break: the watcher’s job is to find a failure before you do

I do not rewrite everything from scratch:
I make what you already have work

I want the same