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

# AI connections

> Connect ChatGPT, Claude, or an OpenAI-compatible model for workflow building, AI nodes, and code writing.

AI connections let sndbox call a model without putting its API key in workflow JSON. A connection selects one provider and model and can be reused by the AI workflow builder, [AI node](/nodes/ai-prompt), and Code editor.

## Add a connection

1. Open **Settings → Connections** and choose **AI provider** under **Add a connection**.
2. Pick the large **ChatGPT**, **Claude**, or **Local / custom** provider card.
3. Choose a model from the dropdown, or select **Custom model ID** and enter the exact ID accepted by the provider.
4. Give the connection a recognizable name and enter its API key.
5. For **Local / custom**, confirm the OpenAI-compatible base URL.
6. Choose **Connect**.

| Choice         | Provider used by sndbox                | Request API                        |
| -------------- | -------------------------------------- | ---------------------------------- |
| ChatGPT        | OpenAI                                 | Responses API                      |
| Claude         | Anthropic                              | Messages API                       |
| Local / custom | Ollama or another endpoint you control | OpenAI-compatible chat completions |

<Note>
  The one-to-five yellow dots are a relative token-usage guide within the model picker. They are not an exact price, context-window size, rate limit, or guarantee of the tokens a request will consume.
</Note>

## Local and custom endpoints

The default local URL is `http://localhost:11434/v1`. Custom remote endpoints must use HTTPS; plain HTTP is accepted only for `localhost`, `127.0.0.1`, or the IPv6 loopback address. Do not embed a username, password, query string, or fragment in the base URL. Put the key in the separate API-key field.

Some local servers ignore authentication. The connection form still requires a non-empty key value; use a harmless placeholder only when the server documentation confirms that authentication is ignored.

## What the connection can power

| Capability          | What the model receives                                       | What changes                                            |
| ------------------- | ------------------------------------------------------------- | ------------------------------------------------------- |
| AI workflow builder | Your instruction and the current workflow graph               | Produces a draft graph for review; it does not run it   |
| AI node             | The configured system instruction, prompt, and mapped context | Pauses the workflow and returns the live model response |
| Code editor         | Your instruction, language, and current editor draft          | Replaces the unsaved draft with a complete code file    |

See [Build workflows with AI](/workflows/ai-builder) for the differences between these surfaces.

## Credentials and privacy

The API key is stored in the operating-system credential vault. Workflow files, exports, and synced definitions keep only the connection ID and model metadata. The underlying key is not exposed to node outputs or ordinary execution logs.

Prompts and relevant workflow or code context are sent to the provider selected by the connection. Review that provider's data policy before sending personal, confidential, or regulated data. A local endpoint keeps the model request on the configured local service only when the endpoint itself is local and does not forward requests elsewhere.

Revoking a connection prevents later AI requests that reference it. Existing workflow definitions retain the unresolved reference until you select another connection.