AI Gateway

The AI Gateway is one OpenAI-compatible endpoint at https://api.dekka.cloud/v1/ai for every model your projects may use. You keep the OpenAI SDK you already have, point its base URL at the gateway and use a gateway key, and switching provider is a change to the model name. Each request is logged with its model, tokens, timing and cost, and you can limit what each project may use with routing rules.

Send your first request

Create a gateway key under AI Gateway, then Keys, as AI keys describes, and save it as DEKKA_AI_KEY. Then send a chat completion with a model from the catalogue, named by provider and model.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.dekka.cloud/v1/ai",
  apiKey: process.env.DEKKA_AI_KEY,
});

const completion = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-5-5",
  messages: [{ role: "user", content: "Summarise this complaint in one sentence." }],
});

console.log(completion.choices[0].message.content);

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.dekka.cloud/v1/ai",
    api_key=os.environ["DEKKA_AI_KEY"],
)

completion = client.chat.completions.create(
    model="anthropic/claude-sonnet-5-5",
    messages=[{"role": "user", "content": "Summarise this complaint in one sentence."}],
)

print(completion.choices[0].message.content)

curl https://api.dekka.cloud/v1/ai/chat/completions \
  -H "Authorization: Bearer $DEKKA_AI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-5-5",
    "messages": [{"role": "user", "content": "Summarise this complaint in one sentence."}]
  }'

The reply is a chat.completion object in OpenAI's shape, with the model that answered as model and the token counts under usage. GET /v1/ai/models lists the models your key may use, in OpenAI's model list format.

Stream the reply

Set stream: true and the gateway answers with server-sent events of chat.completion.chunk, ending in data: [DONE]. The last chunk before [DONE] has no choices and carries the usage, whether or not you set stream_options.include_usage, so read each chunk's choices with care.

const stream = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-5-5",
  messages: [{ role: "user", content: "Write a short welcome message for new customers." }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

If the provider fails after the reply has started, the response is already a 200, so the stream carries an event with error.code and error.message and then [DONE].

See which provider and model answered

Every successful response carries two headers. They matter after a fallback, when the model that answered isn't the one you asked for.

With the OpenAI SDK for TypeScript, withResponse() gives you the headers alongside the reply.

const { data, response } = await client.chat.completions
  .create({ model: "anthropic/claude-sonnet-5-5", messages: [{ role: "user", content: "Hello" }] })
  .withResponse();

console.log(response.headers.get("X-Dekka-Model"), data.choices[0].message.content);

What does the gateway accept?

The gateway takes OpenAI's chat completions request and ignores fields it doesn't know. Messages can have the roles system, developer, user and assistant, and their content must be text, so an image or any other part answers 422 unsupported_content. Tool calling isn't supported yet, and a tools field is ignored.

It passes on max_tokens or max_completion_tokens, temperature, top_p, up to four stop sequences and user. Anthropic needs a maximum, so the gateway sets max_tokens to 1024 when you send none. Claude models from 4.7 on accept only their default sampling settings, so temperature and top_p reach only Claude Haiku 4.5 and Claude Sonnet 4.5 among the Anthropic models.

Where are prompts processed?

OpenAI and Anthropic process every prompt and reply on their own infrastructure in the United States, which is outside Ghana. Dekka keeps each request's model, token counts, timing and cost, and never stores the prompt or the reply.