---
title: LLM API quickstart
url: https://runapi.ai/docs/guides/llm-api/quickstart.md
canonical: https://runapi.ai/docs/guides/llm-api/quickstart
locale: en
---

# LLM API quickstart

RunAPI exposes language models through familiar public protocol shapes.
Point an existing compatible client at `https://runapi.ai`, use a RunAPI
model identifier, and authenticate with a standard API key.

## Choose a protocol

* OpenAI-compatible Chat Completions uses `POST /v1/chat/completions`.
* OpenAI-compatible Responses uses `POST /v1/responses`.
* Anthropic-compatible Messages uses `POST /v1/messages`.
* Gemini-compatible content generation uses
  `/v1beta/models/{model}:generateContent` or `:streamGenerateContent`.

Choose the protocol that best matches your existing client and
application flow. The model catalog identifies which public protocols
each model supports.

## Send a request

This Chat Completions request returns one synchronous response:

```shell
curl "https://runapi.ai/v1/chat/completions" \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4",
    "messages": [{"role": "user", "content": "Summarize why idempotency matters."}]
  }'
```

Keep the request and response in the selected protocol shape. Use the
operation's API Reference for exact fields, tool calling, structured
output, and model-specific constraints.

## Stream output

Set the protocol's streaming option and consume the returned server-sent
events incrementally. For Chat Completions, add `"stream": true`; for
Gemini-compatible calls, use the `:streamGenerateContent` operation.
Close the stream when your client receives the protocol's terminal
event.

## Handle errors and usage

Treat `401` as an authentication failure, `4xx` responses as request or
policy problems, and retry only transient server or network failures.
Read token usage from the final response or terminal stream event in the
selected protocol shape, and keep request identifiers in application
logs for support.