LowRouter provider for Hermes

Adds LowRouter as a provider in Hermes Agent. Pick LowRouter in hermes model, or run hermes chat --provider lowrouter -m auto/mistralai/mistral-large-2512.

LowRouter is an OpenAI-compatible router run from the EU. An auto/<creator>/<model> ID can be served by several European or global hosts, and LowRouter fails over between them. Maintained by the LowRouter team.

Why not a custom endpoint?

You can point a custom endpoint at https://api.lowrouter.ai/v1 and it works. What you get with the plugin instead (Hermes v0.21.5, live API, 2026-10-09):

	Custom endpoint	Plugin
Model list	719 IDs, including every pinned route, embedding and realtime models, and models without tool calling	the 180 auto/ chat models with tool calling, plus your own alias/ models
Context window	read from /v1/models	read from /v1/models. A named provider without this lookup gets Hermes' 256K fallback, or a family default: 262,144 for auto/qwen/qwen3-coder-30b-a3b-instruct, which has 128,000
/usage	nothing	your credit balance in euros

Install

hermes plugins install lowrouter

Get a key at lowrouter.ai/app/keys and put it in ~/.hermes/.env:

LOWROUTER_API_KEY=sk-lr-...

Then pick LowRouter in hermes model, or in ~/.hermes/config.yaml:

model:
  provider: lowrouter
  default: auto/mistralai/mistral-large-2512

No need to set context_length.

Models

The picker shows the entries from /v1/models that Hermes can run:

• auto/<creator>/<model> IDs and your account's alias/<name> models. Pinned routes (<host>/<creator>/<model>) are skipped since the auto/ ID already covers them. An alias has the capabilities of the model it points to.
• Chat models with tool calling. Hermes does everything through tools.
• At least 64,000 tokens of context, Hermes' minimum. If we don't publish a window for a model, it stays in the list and Hermes looks the window up itself.

Tested on 2026-10-09, one turn each with a tool call:

Model	Input tokens	Output tokens
auto/mistralai/mistral-large-2512	37,342	70
auto/mistralai/mistral-small-2603	12,836	81
auto/qwen/qwen3-coder-30b-a3b-instruct	34,453	152
auto/anthropic/claude-haiku-4.5	53,764	226

Hermes falls back to these four when it can't reach the catalog. auto/mistralai/mistral-small-2603 is the auxiliary model (context compression and other background work).

The plugin doesn't touch reasoning effort.

Cost

LowRouter bills in euros. A routed model has several routes (hosts, regions) with different prices, and you pay for the route that served the request. Each response has the amount in usage.cost, with usage.currency: "EUR". Price ranges per model are on lowrouter.ai/models.

Hermes tracks cost in US dollars (/insights, the dashboard). The plugin doesn't convert, so Hermes shows LowRouter sessions with an unknown cost. /usage shows your balance in euros, from /v1/credits.

Security and footprint

• Registers one provider, lowrouter. No tools, hooks, commands or background threads.
• Chat requests go to https://api.lowrouter.ai/v1 (or your model.base_url) through Hermes' own client, with LOWROUTER_API_KEY.
• The plugin's own requests go to the same base URL with the same key, through Hermes' open_credentialed_url: GET /v1/models for the model list and context windows (cached for an hour), and GET /v1/credits on /usage. Without a key the catalog is fetched anonymously; it's public.
• LowRouter forwards your prompts, tool schemas and tool results to the host serving the model. Hosts per model are listed on lowrouter.ai/models.
• No files read or written, no processes, no dependencies beyond Hermes. Hermes reads LOWROUTER_API_KEY; the plugin asks Hermes for it (resolve_api_key_provider_credentials) only to look up context windows.

Compatibility

Hermes 0.21.5 or newer. CI tests against v2026.9.24 (0.21.5) and main daily.

Changelog

0.1.0

• First release: filtered model list with aliases, context windows from the catalog, EUR balance in /usage.

License

Apache-2.0