Hermes Agent setups

Most assistants are a text box pointed at somebody else’s server. They forget the conversation when the tab closes, they cannot be reached from a phone at the weekend, and anything said to them becomes a request to a company the business has no contract with.

The usual response is to stop telling the assistant anything that matters, which leaves it handling trivia. Hermes is the version where the assistant lives on a machine you control, remembers what it was told, and listens without sending the room anywhere.

What Hermes is

Hermes Agent is an open source agent platform with a full time team behind it. It runs as a long lived process, so it holds memory across conversations and keeps a short profile of the person it works for. It builds a library of skills as it goes, and it carries a scheduler, so a nightly report or a weekly check is an ordinary job rather than a separate system.

One gateway process connects it to Telegram, Discord, Slack, WhatsApp, Signal, and a terminal, so a person talks to the same agent from wherever they are. It is not tied to one model provider either. Changing supplier is a change of setting, which keeps the running cost a decision rather than a lock in.

Speech that stays on the machine

Our own setup transcribes speech with Whisper large-v3 running on a local GPU. Audio never leaves the machine to be turned into text, which matters for a business that talks out loud about its customers and its prices. It also means dictation keeps working when the connection does not.

Replies can come back as speech through a hosted voice. That half carries less risk, but it is not nothing: the text of a spoken reply does reach the voice service. Where that line matters, local synthesis is available, at the cost of a heavier install and one more model to maintain.

What the setup involves

The platform installs in an afternoon. The work is everything that makes it survive.

• It runs as a service on the client’s own machine or server, restarted by the system if it stops.
• It is reached through the messaging app the client already uses, and nothing sits on an open port.
• Instructions, memory, and skills are written from the client’s real material.
• Guardrails are written down: what it may touch, what it asks about first, and what it leaves alone.
• Configuration is kept as files with backups, and the restart and rollback are recorded in one place.

The running cost

The hosting is small, and on a machine the client already owns it is close to nothing. Model usage is metered, so a busy agent has a running cost that grows with use, and a local speech model needs a graphics card with room for it.

The larger cost is attention. An agent that can write to a calendar or an inbox needs the same care as any automation that changes records, which is why the guardrails and the written record are where most of the time goes. An assistant is only as trustworthy as the notes that say what it is allowed to do.