# Hermes Agent

[Hermes Agent](https://hermes-agent.nousresearch.com/) is Nous
Research's open-source autonomous agent. It keeps memory and skills
across sessions and can run unattended. Any OpenAI-compatible
endpoint can serve as its model through a custom provider.

Verified against Hermes Agent **v0.21.5** (release `v2026.9.24`). The
config below is re-run against that release every week, including a
tool call round trip.

## Install

Hermes ships through its own installer. It does not build as a
regular Python package:

```bash
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
```

## Configure

Point the model section of `~/.hermes/config.yaml` at LowRouter:

<!-- verify: hermes-config -->
```yaml
model:
  provider: custom
  base_url: https://api.lowrouter.ai/v1
  key_env: LOWROUTER_API_KEY
  default: auto/mistralai/mistral-large-2512
  context_length: 256000
```

`key_env` names the environment variable holding the key, so the key
itself stays out of the file. Export it before starting Hermes:

```bash
export LOWROUTER_API_KEY=sk-lr-...
```

`hermes model` offers the same thing interactively under "Custom
endpoint". Then run a task:

```bash
hermes -z "Summarise what this repository does"
```

`-z` runs one prompt, prints only the final answer and approves
every tool call without asking. `hermes` alone opens the interactive
session.

Three settings people get wrong:

- `base_url` ends in `/v1`. Hermes appends only
  `/chat/completions`; the bare origin gives a `404`, which Hermes
  reports as the provider being "temporarily unavailable" after three
  retries.
- Set `context_length` to the model's real context window, which the
  model browser shows. Without it Hermes guesses, and it plans
  context compression around that guess. It refuses to start with a
  model under 64k, because its system prompt and tool schemas need
  that room.
- Model IDs contain slashes. To switch models, change `default:` in
  the config (and `context_length` with it) rather than typing the ID
  into `/model` inside a session.

## Picking a model

Hermes does everything through tool calls, so pick a model with the
*function-calling* tag. [The model browser filtered to
them](/models?function_calling=yes) lists every one, and the model
needs at least 64k of context.

Each request carries Hermes' system prompt and about 25 tool
schemas. In our runs the first request of a task was already around
25,000 input tokens, before any file contents.
Input price matters more than output price here. The example uses
`auto/mistralai/mistral-large-2512`, served from the EU (256k
context). Models we have run through Hermes' tool loop:

- `auto/qwen/qwen3-coder-30b-a3b-instruct`, a coding model served
  from the EU (128k context).
- `auto/anthropic/claude-haiku-4.5`, a cheaper Anthropic model
  (200k context).

## Recommended setup

- Use a dedicated key. Hermes runs unattended from its scheduler and
  messaging gateways, so set the turn cap (below): until per-key spend
  limits ship, the cap and revoking the key are what end a runaway
  loop.
- Cap the tool loop. `agent.max_turns` in the config is unlimited by
  default; set a number for unattended runs.
- Narrow the tools with `-t` (`--toolsets`) when a task only needs a
  few; fewer tool schemas also means fewer input tokens per request.

## Troubleshooting

- "custom didn't answer after 3 attempts" with `HTTP 404`:
  `base_url` is missing `/v1`.
- `401` on the first request: `LOWROUTER_API_KEY` is not exported in
  the shell that runs Hermes, or `key_env` names a different variable.
- Startup refuses the model for its context size: the model has
  under 64k of context, or `context_length` is set below that.