Hermes Agent
Nous Researchcatalogue onlyTerminal coding agent from Nous Research. Its gate is unusual: a model decides whether a command needs a human, and only shell commands that trip a risk detector are checked at all.
spec
- Maker
- Nous Research
- Kind
- terminal-agent
- Version
- 0.20.4 · checked 2026-08-20
- License
- MIT
- npm
- Not applicable. It is a Python package.
- Repository
- github.com/NousResearch/hermes-agent
getting it
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
The script clones the repository and builds a Python environment managed by uv. A PyPI package exists but its newest release is 0.19.0 while the repository is at 0.20.4, so pip installs a version behind a full minor.
other channels
- PowerShell, per the README's Windows line beside the curl command
how it is built
Its gate works differently from the others here: rather than prompting on every call, a model decides whether a command needs a human, and only shell commands that trip a risk detector are checked at all. Commands run on your machine with no container between them and your files, so that gate is the only thing in the way.
- Language
- Python
- Ships as
- Install script (clone and build)
- What it can enforce
- Gate in the harness
- Out of the box
- Gated by default
- Extends via
- Not recorded yet. We have not opened this one up.
- Embeds via
- Not recorded yet. We have not opened this one up.
- Runtime
- Python
- Licence
- MIT
capabilities it claims
Taken from the project’s own documentation. Claimed, not verified by us. Verifying is what the playground is for.
- tools
- sessions
going deeper
We have not measured this one yet. Opening a harness up means reading its source, sizing every subsystem and recording a real session. That takes a while, and we would rather do three properly than eleven badly.
status: catalogue only · no measured pack