Hermes Agent is Nous Research's self-hosted agent. Same family as the previous page, plus one thing the others do not have: it turns what it did into reusable skills and improves them as it uses them.
In this pageinstall, config, model knobs, sandboxes and gateway, the traps. Not in this pagemy own setup. No ranking against comparable projects either.
| By | Nous Research |
| Open source | MIT, primarily Python |
| Repo | NousResearch/hermes-agent 233,749 stars / 46,868 forks, created 2025-07-22 |
| Version | Latest release v2026.8.18 (v0.20.4, 2026-08-18) |
| Distinctive | A learning loop: writes its own skills, improves them in use, searches its own past sessions, builds a model of you across sessions |
Windows uses PowerShell's iex (irm …/install.ps1); there are also desktop installers
for macOS, Windows and Linux.
Guides like this are full of curl … | bash: download whatever is at a URL and
hand it straight to a shell. If that URL is not the official one,
you have given away the machine — voluntarily and without a prompt.
And this project has several similarly named sites: searching on 2026-08-21
also surfaced hermes-agent.org and hermes-ai.net alongside the official site.
This page makes no claim about what those are — but you must establish which one is official.
A check that works for any open-source project: find the project's GitHub repository and read the homepage / website field on it. That is filled in by the maintainers. Trust that, not the search ranking.
Prerequisites are light: Git only (plus curl and xz-utils on Linux).
Python 3.11, Node 22, ripgrep and ffmpeg are pulled in by the installer via uv —
you do not install them yourself.
It lands in ~/.hermes/hermes-agent/ with the binary at ~/.local/bin/hermes;
root installs go under /usr/local/.
| File | Holds |
|---|---|
~/.hermes/config.yaml | Settings |
~/.hermes/.env | Secrets |
hermes config set routes automatically: API keys to .env,
everything else to the YAML. That split is what makes
backups straightforward — back up config freely, exclude secrets.
The model configuration is the three knobs from the spine page:
And every model slot uses those three — main, auxiliary, compression, fallback.
base_url points at any OpenAI-compatible endpoint, authenticating with OPENAI_API_KEY.
Providers can be Nous Portal (hermes setup --portal, OAuth, no manual key), OpenRouter, OpenAI,
or your own endpoint; hermes model switches interactively.
Seven execution backends — deciding where its commands actually run:
| Backend | Means |
|---|---|
local | On this machine. Most convenient, least isolated |
Docker | In a container; your machine stays protected |
SSH | On a remote machine — the bridge from local to cloud, see page 17 |
Modal / Daytona | Serverless, hibernates when idle, costing almost nothing unused |
Singularity / Vercel Sandbox | Two more container/sandbox environments |
The messaging gateway decides where you talk to it from: Telegram, Discord, Slack,
WhatsApp, Signal, Email, Home Assistant. Start with hermes gateway setup;
hermes gateway install registers it as a system service (systemd on Linux, launchd on macOS).
Its default policy deserves credit: deny everyone who is not allow-listed or DM-paired. Details on page 15.
Plenty of agents have memory. Hermes takes one more step: it records how something was done, turns that into a reusable skill, and improves it next time. Alongside that sit cross-session retrieval and a running model of the user. Skills are portable and shareable against an open skills standard.
An auto-generated skill is the model's summary of its own past behaviour. It can freeze one lucky success into "this is how we always do it".
So look through what it has written for itself now and then and delete the wrong ones — like project memory, this is a thing to prune, not accumulate.
| Command | Does |
|---|---|
hermes / hermes --tui | Start; the latter is the newer interface |
hermes --continue | Resume the last session |
hermes model | Switch model |
hermes tools | Enable/disable tools — see page 16 |
hermes doctor | Run this first when something is wrong |
hermes claw migrate | Move in from OpenClaw — next page |
NousResearch/hermes-agent, Python, MIT,
233,749 stars / 46,868 forks, created 2025-07-22, latest release v2026.8.18
(v0.20.4, 2026-08-18). GitHub API, checked 2026-08-21.install.sh / install.ps1; only prerequisite is Git
(plus curl and xz-utils on Linux); Python 3.11 / Node 22 / ripgrep / ffmpeg installed
by the script via uv; lands in ~/.hermes/hermes-agent/ with ~/.local/bin/hermes.
Official docs, checked 2026-08-21.~/.hermes/config.yaml + ~/.hermes/.env; every model slot uses
provider / model / base_url, with base_url officially described as
"points at a custom OpenAI-compatible endpoint, uses OPENAI_API_KEY for auth".hermes-agent.org and
hermes-ai.net besides the official site. This page makes no claim about their nature —
it only insists you confirm the official domain before curl | bash,
ideally via the repository's homepage field.