nodecode-lowrouter

LowRouter: its EU-hosted, OpenAI-compatible routes as a Nodecode provider, filtered to the ones a turn can run on, with each route's real context window and your credit balance in euros. A Nodecode rewrite of carboniferio/hermes-lowrouter-provider by carboniferio.

Get it

Tell the agent "add the lowrouter cell", or run nodecode add nc://lowrouter.

Set it up

Make a key at https://lowrouter.ai/app/keys and save it with /connect (pick LowRouter), or leave it in LOWROUTER_API_KEY. Without a key the picker still fills — /v1/models answers anonymously — and the balance waits for one.

Optional, and rarely needed — the section is on unless it says "enabled": false:

{
  "lowrouter": {
    // where the API is served, appended /chat/completions, and where
    // /models and /credits are read. LowRouter's own address; a relay's
    // goes here.
    "base_url": "https://api.lowrouter.ai/v1"
  }
}

Use it

/models lists the routes an auto/ or alias/ id names, and a turn on one goes to https://api.lowrouter.ai/v1/chat/completions with the key as its bearer.

The picker is the filtered catalog, not the raw /v1/models list: an auto/<creator>/<model> id (the one that fails over between European and global hosts) or the account's own alias/<name>, a chat model, one that takes tool calls, with at least 64,000 tokens of context. Pinned <host>/<creator>/<model> routes are left out because the auto/ id already covers them, as are embeddings and realtime models. Each route carries the window the catalog publishes for it, so compression runs when the model's own window fills rather than at one number assumed for the whole endpoint.

alias/coder (or whichever alias you made) is usually the good default: an alias picks a model once, for every session.

/lowrouter shows your credit balance in the currency LowRouter bills in (euros, for an EU account) and how many routes the picker holds; /lowrouter models reads the live catalog, lists the routes that take tool calls and takes them into the picker at once.

LowRouter bills in euros and reports the amount of each response in usage.cost. Nodecode counts cost in US dollars, so a LowRouter session reads with an unknown cost, and the balance above is the amount left.

What this rewrite carries, and what it leaves to Nodecode

The plugin registers one Hermes model provider: a base, a key variable, a filter over /v1/models, context windows read from the same answer and cached an hour per key and base, /v1/credits on /usage, four fallback models, and an auxiliary model.

The plugin's model filter is the cell's, unchanged: auto/ and alias/ ids, the chat modality, tools among the supported parameters, and a 64,000-token floor with an unpublished window passing.
The window lookup. Nodecode keeps a catalog model's window beside the model, so the live answer's context_length becomes the row's window; there is no per-key cache to keep, because the worker refetches the catalog hourly and a restart fetches again.
/usage. Nodecode's cost accounting is a dashboard's, not a command's; the balance is /lowrouter here, read from the same /v1/credits answer and formatted as the plugin formats it.
The auxiliary model. The plugin points Hermes' background work (context compression, summaries) at auto/mistralai/mistral-small-2603; Nodecode's is general.auxiliary_model, which the README here can only suggest.
The fallbacks. The plugin's four models stay as the picker's rows until the live catalog answers, and again whenever it does not.
Nothing is converted. The plugin does not convert EUR to USD, and neither does this cell.

Test

From this repository's root:

.github/test-cell.sh providers/nodecode-lowrouter

Every request is a stubbed nlk:http and every key variable a stubbed nle::credential-env: nothing reaches LowRouter.

Apache-2.0 licensed.