Hermes, From Chat to Agent
Hermes is an open-source AI agent I can message on Telegram. I connected it to my tools because I wanted help with real work, not another chat window.
This is Part 4 of the AI Agents series. If you’re new to agents, start with why you’re overpaying for AI tools, then stop treating agents like magic.
I have spent enough time using AI to write code to know the pattern. I ask a question, get an answer, and then carry on with the rest of the work myself.
When I watched Dreams of Code set up Clawdbot, I liked that it showed an assistant doing more than talking. It lived on a server, answered through Telegram, and could use tools to handle ordinary digital work.
I already had Hermes, an open-source AI agent from Nous Research, running on my server. I was mostly using it as a technical assistant. I wanted to see what changed when I connected it to more of my day.
What Hermes is
Hermes is not the language model itself. It is the agent around the model: the part that receives a request, works out what it needs, and uses tools to carry it out.
In my setup:
- Hermes runs on my server and keeps its configuration, skills and memory.
- Telegram is where I message it from my phone.
- n8n connects it to services such as Gmail, Google Calendar and Google Tasks.
- A language model provides the reasoning, while tools provide access to information and actions.
The model can change. The agent’s tools and context are what make it useful for a particular job.
The first real test
I asked Hermes to help me get on top of my work. For that, it needed to see more than the current conversation. I tested its connections to Gmail, Calendar and Tasks.
The Google authorisations had expired. The workflows were still there, but the first tests failed until I reconnected the accounts. After that, Gmail returned messages and Calendar returned an event. The default Google Tasks list returned one item, already completed.
That was a useful result. It showed the connection worked, and it showed that the list I had connected was not a complete picture of my commitments. Hermes could inspect what I had captured. It could not find work I had never put in the list.
Useful, not unattended
The connection is only the start. I want Hermes to help me gather information, spot what needs attention and choose a sensible next step. I still want to approve actions that affect other people or change my commitments.
That is also why I liked the Clawdbot video. It did not stop at an installation screen. It showed a working assistant, then spent time on the parts that can go wrong. An agent with access to email or a server needs boundaries as much as it needs tools.
Hermes has not solved my productivity. It has given me a system I can keep improving, one connection and one workflow at a time.
If you are trying an agent, start with a task you actually want off your plate. Connect only the information it needs, test what it can see, and keep consequential actions under your control.
That is the version of AI assistance I wanted to try.