OpenAI compatibility

Any client that speaks the OpenAI API can call Noket. Set the base URL to https://api.noket.dev/v1, use a Noket API key, and pass a Noket model id as model.

Chat completions

curl https://api.noket.dev/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "alice/llama-3.1-8b",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

The response is the model provider’s own response body. Noket passes it through rather than wrapping it, so the fields your client reads — choices, usage, tool calls, and so on — are the ones that model produces.

Point an SDK at Noket

from openai import OpenAI

client = OpenAI(
    base_url="https://api.noket.dev/v1",
    api_key="YOUR_API_KEY",
)

response = client.chat.completions.create(
    model="alice/llama-3.1-8b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

The openai SDK reads OPENAI_BASE_URL and OPENAI_API_KEY from the environment, so an existing application can switch by setting both and changing only the model id.

What Noket checks

Noket validates only what it needs to route the request:

• model must be a non-empty string.
• stream, when present, must be a boolean.
• messages must be a non-empty array for chat completions.
• prompt is required for legacy completions.
• input is required for the Responses API.

Everything else is forwarded to the model unchanged. If the provider behind the model does not accept a field, the provider rejects the request and Noket returns 400 with upstream_invalid_request.

For safety, Noket removes fields that would let a client choose a provider or supply its own credentials — provider, providers, fallback, fallbacks, routing, route, virtual_key, api_key, base_url, headers, extra_fields, _bifrost, bifrost, and headers prefixed x-bf- or x-bifrost-. The model’s stored configuration decides those, not the request.

Streaming

Send stream: true as usual. The response is a server-sent event stream of the provider’s chunks.

For chat completions, Noket always sets stream_options.include_usage on the forwarded request, so the last chunk carries usage even if the client left it off. A model configured without streaming support returns 400 with stream_not_supported.

Response contents

Responses never expose internal routing. Noket:

• replaces the internal model id with the public id;
• removes provider metadata (extra_fields);
• removes usage.cost from the payload — Noket’s cost accounting stays in the console;
• sends Cache-Control: no-store.

List models

curl https://api.noket.dev/v1/models \
  -H "Authorization: Bearer YOUR_API_KEY"

{
  "object": "list",
  "data": [
    {
      "id": "alice/llama-3.1-8b",
      "object": "model",
      "created": 1730000000,
      "owned_by": "alice"
    }
  ]
}

The list contains the models this key can call right now: the account’s own active models and the models shared with it. Draft and paused models and expired access are not listed. owned_by is the owner’s username, or noket when the account has none. The same list is available in the Anthropic format.

Next

• Anthropic compatibility
• Errors and limits
• Examples