Bot fleet

What this actually is

One machine runs Hermes Agent, a self-hosted open-source agent runtime. On top of it sit seven profiles, each a separate assistant with a narrow job, its own skills, its own memory and its own standing instructions about what it may and may not do. They are not personas painted on one assistant. They are separate configurations with separate state.

The interesting work in a setup like this is not the runtime. It is deciding where one assistant's authority stops, what it is allowed to touch without asking, what it must escalate, and how seven of them share a machine without stepping on each other.

The runtime, named

It is Hermes Agent, an MIT-licensed agent runtime from Nous Research. This page used to call it "a third-party open-source agent runtime" and stop there, which told a visitor nothing except that something was being withheld. Naming what you run is disclosure, not promotion, so it is named.

What it supplies is the part I did not build: the agent loop and its tool calls, the profile system that lets one machine run several assistants with separate state, the gateway that multiplexes them, the scheduler behind their standing jobs, the skills loader, and the message tool the profiles use to reach each other. That list is upstream's work, and it is most of what makes a fleet possible at all.

The copy running here is v0.20.5 (2026.8.19), installed from git and sitting on upstream's main. I track it by pulling rather than patching. An earlier checkout of mine did carry local commits, and keeping them alive against a repository that lands changes daily cost more than the patches were worth, so they went. The tree is clean and zero commits ahead of upstream, which is this page's authorship claim stated as a number instead of a promise.

What is mine and what is not

The checkout is the line. Everything below it belongs to the people who wrote Hermes Agent: profile support, the multiplex, the scheduler and the agent-to-agent message tool are all upstream features. I did not build an agent framework and this page is not going to imply that I did. I have filed six bug fixes into it, which is a different claim and is recorded, outcomes and all, on the upstream page.

What is mine is everything above that line: the seven profiles, their roles and boundaries, the skills each one carries, the scheduled work, the routing that decides which profile answers what, and the operating rules that keep them from acting outside their scope.

Why there is nothing to open

It is a configuration, not a product

Hermes Agent itself is public and anybody can install it. What is not public is my configuration of it. There is no installer for that, no repository and no instance anybody else can run. It is a working setup on one machine, and the register says "Private" because that is exactly what it is.

Not version controlled

The configuration tree is not a git repository at all, so nothing has ever been pushed anywhere. That is not caution after the fact; it has never had a remote to push to.

It holds live credentials

Each profile keeps its own authentication next to its configuration. Publishing the tree would mean publishing those, so no part of it is reproduced here, including key names.

No roster on this page

Seven is the fact the register carries. What each of the seven does is personal, and a public page listing them would disclose more about my week than about the engineering.