Hermes Agent on a VPS: Three Telegram Bots for Obsidian, Email, and Server Audits

⚡ The Gist in 5 Seconds

• The author of the Timeweb Cloud blog deployed the open-source Hermes Agent on a dedicated server and split it into three profiles, each with its own Telegram bot.
• The bots handle three tasks: a Zettelkasten-based knowledge base in Obsidian, server audits for the ServeHub-2 project, and triaging three email inboxes.
• Access to internal services goes only through the 10.8.0.0/24 virtual network — no open ports and no port forwarding on the router.

🔍 What Happened

An author going by ITsuperiorRF published a case study in the Timeweb Cloud blog on Habr: he rented an additional server at Timeweb Cloud, connected it to his ServeHub-2 network, and deployed Hermes Agent there — an open-source AI agent that executes tasks via tools and works with files and the command line. The agent is split into three profiles, each with its own Telegram bot: one maintains the Obsidian knowledge base, the second audits servers, and the third keeps an eye on email. The main model is DeepSeek V4 Flash via API.

The mechanics sound simple in words but are complex in the details. Hermes' capabilities are extended through skills, and the profiles separate tasks and settings. Plus, there's built-in self-improvement: the agent stores useful takeaways from past tasks in long-term memory and refines its own skills. The author specifically clarifies that the base model isn't retrained — it's only about memory and working scenarios. The network is set up like this: Hermes runs on a dedicated server and reaches the remote VPS and home server only within 10.8.0.0/24. No ports are exposed outward, no router port forwarding is needed — but the device must be on this network, and access to it must be configured in advance.

🧠 Our Take

It looks like this case grew out of everyday pain, not out of a desire to play with an agent. The author already has monitoring and alerting, but they react to pre-defined events and don't explain the full picture: a container goes down — a notification arrives, and figuring out what actually happened is a manual job. Obsidian needs its own scripts, email needs its own tools, and the result is a zoo that's inconvenient to manage. An agent with tools here is an attempt to bring everything into a single point of entry, and the entry point chosen is the cheapest possible: Telegram.

What does this change, our assessment: this isn't news about a model, it's news about an interface to infrastructure. One VPS plus token costs versus several subscriptions and hand-written scripts — the arithmetic favors the agent as long as the tasks are routine. But the author honestly rules out local models on quality grounds, and that's the key fork: the privacy of notes and email moves to a cloud API. I'd bet that this exact point, not the price, will be the main dealbreaker for most people.

The winners are VPS and API providers — the case is precisely about how to sell one more server for an agent. The losers are niche SaaS wrappers that do one thing: "chat with your notes," "chat with your logs." But the main risk isn't the market. A Telegram bot with ssh and email access has a huge blast radius, and the article says almost nothing about that.

⚠️ What We Don't Know

There isn't a single number: how many tokens are spent per month, what it costs, what the latency is on a typical task, how access to the bots is restricted (who can even message them and get the reins to the servers), where secrets are stored, and what gets written to the agent's action log. The DeepSeek V4 Flash model is named without pricing and without comparison to alternatives — you can't verify its behavior on long multi-step tasks from a single post. And this is exactly one voice: a personal case study published in a hosting provider's corporate blog, with no independent measurements.

Forecast: over the next quarter, expect a wave of similar "agent as a single entry point into self-hosted" case studies — with published profile configs and Telegram wrappers. Easy to verify: will ready-made repositories or images with a preconfigured Hermes profile appear, rather than just text guides?

Disclaimer: a single source — the author's personal case study in a corporate blog, not independently confirmed.

Source: Habr / Timeweb Cloud