> ## Documentation Index
> Fetch the complete documentation index at: https://docs.nora.my/llms.txt
> Use this file to discover all available pages before exploring further.

# Internal resources

> Register in-house LLM endpoints and SQL databases once, then reference them by name from any block.

**Internal resources** is where you register in-house LLM endpoints and SQL databases once. After that, any block can reference them **by name**, with no pasting connection strings per block.

The list shows each resource's **type** (LLM / DB), **name**, and **address**. Edit or delete from the row.

## Adding an endpoint

**Settings → Internal resources → Add → Endpoint**:

* **Name**: what blocks will call it (e.g. `GPU box A`).
* **Address**: an OpenAI-compatible base URL (e.g. `http://localhost:11434/v1`).
* **Purpose** (optional): what the endpoint serves: Inference, Embedding, Reranker, Retrieval, OCR, Voice, or Image.
* **Custom headers** (optional): HTTP headers sent with every request. Put authentication here (e.g. `Authorization: Bearer sk-...`). Blank lines are dropped on save.

**Test connection** calls `GET /v1/models` (falling back to `/api/tags` and `/models`) against the address with your headers. It runs from the server, so browser CORS doesn't matter.

### Set up a GPU box

For an empty GPU server, **Set up a GPU box** installs a sidecar in one click: Nora issues a token, and you run the printed install command on the GPU machine. The endpoint registers itself when the sidecar comes up.

## Adding a database

**Settings → Internal resources → Add → Database**:

* **Name**: e.g. `prod-orders`.
* **Database URL**: `postgres://…`, `mysql://…`, or `mariadb://…`.

**Test connection** opens a short-lived pool and runs `SELECT 1`. Failures aren't cached, so retesting after a fix is safe.

## Using a resource in a block

* **SQL tool**: pick the registered database from the **Database** dropdown; queries run over that connection (see [Tool block](/build/workflow/tool-block)).
* **LLM endpoints**: selectable wherever an in-house model can run: [Ask AI on-prem routing](/settings/engine/ask-ai) and the [Loop's on-prem runtime](/settings/engine/loop).