Chat Completions

Send a list of messages, name a model, get a reply in the OpenAI chat-completions shape. The model ID carries the privacy level.

Endpoint

POST /v1/chat/completions accepts OpenAI-compatible requests. Unknown fields at gateway-owned objects are dropped for compatibility. A control the model cannot honor returns 400 invalid_request before any usage is reserved.

Request

Only messages is required.

Field	Type	Notes
`messages`	array	Required. An empty array is rejected.
`model`	string	Logical ID such as `openai/gpt-5.5`. Add `:confidential` for the attested variant.
`temperature`	number	Forwarded when set.
`top_p`	number	Forwarded when set.
`max_tokens`	integer	Cap on generated tokens, used when `max_completion_tokens` is unset.
`max_completion_tokens`	integer	Preferred cap.
`stream`	boolean	Server-Sent Events. See Streaming.
`stream_options.include_usage`	boolean	Trailing usage chunk on the stream.
`reasoning`	object	See Reasoning.
`reasoning_effort`	string	Flat OpenAI-style effort: `minimal`, `low`, `medium`, `high`. `reasoning.effort` wins when both are present.
`web_search`	boolean	See Web Search.
`web_fetch`	boolean	See Web Fetch.
`tools`	array	Function tools plus the aliases `{"type":"web_search"}` and `{"type":"web_fetch"}`. See Tool Calling.
`tool_choice`	string or object	`auto`, `none`, `required`, or a named function.
`response_format`	object	JSON mode or JSON Schema, on models advertising `json_mode` or `structured_outputs`.
`stop`, `frequency_penalty`, `presence_penalty`, `parallel_tool_calls`	various	Accepted on models whose catalog entry advertises the control; otherwise `400`.
`user`	string	Opaque end-user identifier, forwarded where supported.

n: 1 and store: false are accepted as no-ops. n above 1, store: true, and audio output are rejected. The same feature sent twice (for example web_search as a field and as a tool) is rejected. A model’s supported optional controls are in supported_parameters on its catalog entry. The supports query can filter for controls such as json_mode; structured_outputs qualifies response_format rather than naming a top-level request field. See Model Discovery.

Messages

Each message has role and content. Roles: system, user, assistant, tool, developer. A tool message pairs with a tool_call_id. content is a string or an array of parts for multimodal input.

Suffixes on the Model ID

Clients that can only set model can carry a few controls in the ID: z-ai/glm-5.2?reasoning_effort=high, moonshotai/kimi-k3?web_search=auto. Suffix and body must agree when both are present. The suffix is stripped before routing. See Model ID Convention.

Example

curl https://app.trynito.ai/v1/chat/completions \
  -H "Authorization: Bearer $NITO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "messages": [
      {"role": "system", "content": "You are concise."},
      {"role": "user", "content": "Explain a gateway in one sentence."}
    ]
  }'

from openai import OpenAI

client = OpenAI(
    base_url="https://app.trynito.ai/v1",
    api_key="YOUR_NITO_API_KEY",
)

resp = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[
        {"role": "system", "content": "You are concise."},
        {"role": "user", "content": "Explain a gateway in one sentence."},
    ],
)

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

import OpenAI from "openai";

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

const resp = await client.chat.completions.create({
  model: "moonshotai/kimi-k3",
  messages: [
    { role: "system", content: "You are concise." },
    { role: "user", content: "Explain a gateway in one sentence." },
  ],
});

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

For attested hardware, use a :confidential variant, for example "model": "moonshotai/kimi-k3:confidential". GET /v1/models lists their :confidential IDs. See Privacy Levels.

Response

{
  "id": "...",
  "object": "chat.completion",
  "created": 1751299200,
  "model": "moonshotai/kimi-k3",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "A gateway is a single entry point that routes your request to the right backend." },
      "finish_reason": "stop"
    }
  ],
  "system_fingerprint": null,
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 17,
    "total_tokens": 41
  }
}

Exactly one choice. Read choices[0].
finish_reason: stop, length, tool_calls, or error.
system_fingerprint is always null.
usage: prompt_tokens, completion_tokens, total_tokens, plus reasoning_tokens, server_tool_use, cached-token details, and customer_cost when relevant. See Check Usage.

Errors

Validation problems return 400 with {"error": {"message", "code", "param"}}. Non-POST requests return 405 with Allow. All responses carry Cache-Control: no-store. See the error reference.

Other Text Endpoints

POST /v1/messages for Anthropic-shaped clients. See the Anthropic drop-in.
POST /v1/responses for stateless Responses-API clients. See the Responses reference.

Tool Calling

Function calls.

Reasoning

Think first.

Streaming

Token by token.

Privacy Levels

The four levels.