nc://empiriolabs

EmpirioLabs AI: its live chat catalog as a Nodecode provider, with tool calling, image input and reasoning declared per model. A Nodecode rewrite of EmpirioLabs-AI/hermes-empiriolabs by EmpirioLabs-AI.

EmpirioLabs AI serves open models on infrastructure of its own behind the OpenAI chat wire at https://api.empiriolabs.ai/v1, and publishes its catalog at GET /v1/models. With this cell, /connect offers EmpirioLabs AI, /models lists its chat models, and a turn on empiriolabs/<model> goes to that service with your key.

The catalog is where the details come from, as the original had it:

• The model list is the catalog's chat models that take tool calls. A turn always offers tools and reads text back, so a row without POST /v1/chat/completions in supported_endpoints, without function_calling in features, or marked deprecated is left out. A base that answers a plain OpenAI-style list (an EMPIRIOLABS_BASE_URL of your own) has none of those fields, and is taken whole.
• Each model's details are declared, not guessed. The context window (context_length or context_window), image input (input_modalities or an image/vision feature), reasoning (capabilities.reasoning or a reasoning feature), the reasoning_effort levels the model lists under supported_parameters, and its input/output price per token all reach /models, the reasoning picker and the turn.
• reasoning_effort is sent only when the model takes it. With a level chosen and the model declaring levels, the round asks for the chosen one, or the nearest level below it — never a dearer one — and the weakest when the chosen one is below all of them. A model that declares none is sent nothing, so the service's own default applies.
• The two ids the original falls back to (glm-5-3, qwen3-7-flash) are listed before the catalog has answered, so the picker is never empty.

Get it

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

Set it up

Make a key at https://platform.empiriolabs.ai/dashboard/api-keys. The picker asks for it when you choose EmpirioLabs AI; a key saved with /connect wins over the environment. EMPIRIOLABS_API_KEY works the same way, as the original read it.

The base is optional, for a relay in front of the service:

{
  "empiriolabs": {
    // where the service is served; the chat lane appends /chat/completions
    "base_url": "https://api.empiriolabs.ai/v1"
  }
}

Use it

Run /models, choose EmpirioLabs AI and pick a model from the catalog. /think offers the levels that model declares. Then talk as usual: prompts, tool calls and tool results go to api.empiriolabs.ai (or your base_url) with your key, and are billed to the account that owns it.

Not ported

• The original's one-hour in-memory catalog cache with its back-off after a failed fetch. The picker's own listing cache already holds a fetched listing for ten minutes and a failure for one, and this cell adds no second copy; its catalog is refreshed by the listing the picker and /connect already make.
• The original's auxiliary-model pick (the lowest-priced featured chat model with tool calling and a 200K context). Nodecode keeps one auxiliary target for the whole image — /models aux — rather than one per provider, so the choice is yours there, with the price of every model in the catalog row.
• The original's gzip handling of the catalog body: nlk:http negotiates and decodes the response itself.

Test

From this repository's root:

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

Every catalog GET is a stubbed nlk:http, every key variable a stubbed nle::credential-env, every round a stubbed dex:post: nothing reaches the service.

MIT licensed.