OpenAI Migration

If you already use the OpenAI SDK, keep the client and change the API key, base URL, and model.

#What changes

Use the same chat.completions.create request shape where supported. The API key authenticates with Elektric, the base URL becomes https://elektric.ai/v1, and the model selects Grid or Wire.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.ELEKTRIC_API_KEY,
  baseURL: "https://elektric.ai/v1",
});

const response = await client.chat.completions.create({
  model: "elektric-grid",
  messages: [{ role: "user", content: "Hello" }],
});

#Use Grid

Set model: "elektric-grid" and Elektric selects an eligible model for the request.

#Use Wire

Wire sends the request to the exact public model ID you provide instead of routing it through Grid. Use model: "<public-model-id>" and discover current IDs with the Models page or authenticated GET /v1/models.

#What stays the same

• chat.completions.create(...) and the standard message roles remain available.
• Chat streaming is supported through the OpenAI-compatible SSE response.
• Function tools are supported for non-streaming tool turns.
• Structured output is supported with response_format using type: "json_schema" on non-streaming requests.

#Compatibility limits

Elektric supports a focused OpenAI-compatible surface, not every OpenAI endpoint or option.

• The supported endpoint is POST /v1/chat/completions; use Elektric’s native APIs for other resources.
• Unknown request fields are rejected. n supports only 1; use either max_tokens or max_completion_tokens, not both.
• Streaming function tools and streaming structured output are not supported. Tool turns must be non-streaming, and structured output must use the supported JSON-schema form.
• Wire capability and parameter support varies by selected model. Images, documents, web search, reasoning, tools, and output limits are model-specific; check GET /v1/models.

#Errors and debugging

Authentication failures return 401 authentication_error. Invalid requests and unsupported fields return 400 invalid_request_error; an unavailable named model returns 400 model_selection_not_supported. Responses include a Request ID; inspect it with Usage logs when debugging. See Errors and limits →

#Next steps

Next: learn how Grid routes, use Wire for exact model selection, or browse available models.