nc://scaleway

Scaleway Generative APIs: Scaleway's European, OpenAI-compatible inference as a Nodecode provider, on a curated tool-calling model list with each model's real context window and price. A Nodecode rewrite of scaleway/hermes-scaleway-provider by scaleway.

Get it

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

Set it up

Make an IAM API key at https://console.scaleway.com/iam/api-keys; the IAM principal needs the GenerativeApisModelAccess permission set. Its secret is the key. Save it with /connect (pick Scaleway), or leave SCW_SECRET_KEY in the environment — the same variable the Scaleway CLI and SDKs export, so an existing CLI setup works as it is. A key /connect saved answers first; the variable SCALEWAY_API_KEY is read too, when SCW_SECRET_KEY is unset.

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

{
  "scaleway": {
    // where the API is served, appended /chat/completions. A relay's
    // address goes here, or SCALEWAY_BASE_URL in the environment, which
    // sets the default.
    "base_url": "https://api.scaleway.ai/v1"
  }
}

Use it

/models lists Scaleway's models, and a turn on one goes to https://api.scaleway.ai/v1/chat/completions with the key as its bearer.

The picker is the curated list, not the raw /v1/models catalog: the serverless models that drive an agent over chat-completions, each with the context window, output limit, price and image modality models.dev publishes for it, so compression runs when the model's own window fills rather than at one window assumed for the whole endpoint. deepseek-v4.1-flash-1017 is a good default. gpt-oss-120b is deliberately absent — Scaleway serves its tool-calling on the beta Responses API — as are the embedding, speech and transcription models.

For the plugin's cheap auxiliary tier, point the auxiliary model at mistral-small-3.2-24b-instruct-2506, the cheapest tool-calling vision-capable model on the platform.

What this rewrite carries, and what it leaves to Nodecode

The plugin contributes a provider profile: a base, an alias, a key variable, a curated fallback_models list, a models_dev_id, and an auxiliary model. Nodecode's catalog already speaks models.dev, so the cell adds only what a profile says that a catalog cannot: the curated list itself, the recorded numbers under it for a home that has never fetched models.dev, and the ambient SCW_SECRET_KEY.

• The alias. Hermes answers to --provider scw; Nodecode names a provider once, so a turn names scaleway/<model>.
• Model metadata. Nodecode reads models.dev at runtime, so a model the core has fetched keeps models.dev's own numbers; models.json here is the row the plugin records, for a curated id the live catalog does not carry yet.
• The setup default. The plugin's setup picks deepseek-v4.1-flash-1017 first. Pick it once with /model; there is no per-provider default to set from a cell.
• Never auto-detected. Hermes refuses to infer the provider from the ambient key. Nodecode may infer it when no provider is configured and SCW_SECRET_KEY is the only key in the environment; naming the provider in providers.scaleway or picking a model settles it.

Test

From this repository's root:

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

Every store is a temp auth.json, every key variable a stubbed nle::credential-env: nothing reaches Scaleway.

MIT licensed.