Text generation

This section has moved to Streaming.

Use the Responses resource when you want the MKA1 API to return text. Start with a plain string for simple prompts. Use message items when you need explicit roles or conversation state. Later sections cover sampling and length limits, reasoning effort, and the request fields the gateway drops. Pausing a background response has its own section in background responses.

Send a simple prompt

Pass a string in input for a single-turn request. The response includes generated text in output_text. If a guardrail blocks the request, output_text is absent and the output carries a refusal instead; see Guardrails.

import { SDK } from '@meetkai/mka1';
import type { ResponseObject } from '@meetkai/mka1/models/components';

const sdk = new SDK({ bearerAuth: 'Bearer <mka1-api-key>' });

const result = await sdk.llm.responses.create({
  xOnBehalfOf: '<end-user-id>', // optional — attribute the request to one of your end users
  responsesCreateRequest: {
    model: 'auto',
    input: 'Write a one-sentence summary of the MKA1 API.',
  },
});

import OpenAI from 'openai';

const openai = new OpenAI({
  apiKey: '<mka1-api-key>',
  baseURL: 'https://apigw.mka1.com/api/v1/llm/',
  defaultHeaders: { 'X-On-Behalf-Of': '<end-user-id>' },
});

const response = await openai.responses.create({
  model: 'auto',
  input: 'Write a one-sentence summary of the MKA1 API.',
  stream: false,
});

from meetkai_mka1 import SDK

sdk = SDK(bearer_auth="Bearer <mka1-api-key>")

res = sdk.llm.responses.create(
    model="auto",
    input="Write a one-sentence summary of the MKA1 API.",
    x_on_behalf_of="<end-user-id>",  # optional — attribute the request to one of your end users
)

using MeetKai.MKA1;
using MeetKai.MKA1.Types.Components;

var sdk = new SDK(bearerAuth: "Bearer <mka1-api-key>");

var res = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Input = ResponsesCreateRequestInput.CreateStr(
        "Write a one-sentence summary of the MKA1 API."),
});

mka1 llm responses create \
  --model auto \
  --input '"Write a one-sentence summary of the MKA1 API."' \
  -H 'X-On-Behalf-Of: <end-user-id>'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "input": "Write a one-sentence summary of the MKA1 API."
  }'

If you are not acting for an end user, omit X-On-Behalf-Of.

Add instructions

Use instructions to define behavior before the model sees the user input. Keep instructions short and specific.

const result = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    instructions: 'You are a support assistant. Reply in plain English. Keep answers under 80 words.',
    input: 'Explain what embeddings are used for.',
  },
});

const response = await openai.responses.create({
  model: 'auto',
  instructions: 'You are a support assistant. Reply in plain English. Keep answers under 80 words.',
  input: 'Explain what embeddings are used for.',
  stream: false,
});

res = sdk.llm.responses.create(
    model="auto",
    instructions="You are a support assistant. Reply in plain English. Keep answers under 80 words.",
    input="Explain what embeddings are used for.",
)

var res = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Instructions = "You are a support assistant. Reply in plain English. Keep answers under 80 words.",
    Input = ResponsesCreateRequestInput.CreateStr("Explain what embeddings are used for."),
});

mka1 llm responses create \
  --model auto \
  --instructions 'You are a support assistant. Reply in plain English. Keep answers under 80 words.' \
  --input '"Explain what embeddings are used for."'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "instructions": "You are a support assistant. Reply in plain English. Keep answers under 80 words.",
    "input": "Explain what embeddings are used for."
  }'

Send structured messages

Use an array of message items in input when you want explicit roles. Each message item uses type, role, and content.

const result = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    input: [
      { type: 'message', role: 'developer', content: 'Answer as a technical writer. Keep the reply concise.' },
      { type: 'message', role: 'user', content: 'Draft a short product update about faster response times.' },
    ],
  },
});

const response = await openai.responses.create({
  model: 'auto',
  input: [
    { type: 'message', role: 'developer', content: 'Answer as a technical writer. Keep the reply concise.' },
    { type: 'message', role: 'user', content: 'Draft a short product update about faster response times.' },
  ],
  stream: false,
});

res = sdk.llm.responses.create(
    model="auto",
    input=[
        {"type": "message", "role": "developer", "content": "Answer as a technical writer. Keep the reply concise."},
        {"type": "message", "role": "user", "content": "Draft a short product update about faster response times."},
    ],
)

var res = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Input = ResponsesCreateRequestInput.CreateArrayOfItem(new List<Item>
    {
        Item.CreateInputMessage(new InputMessage()
        {
            Role = InputMessageRole.Developer,
            Content = InputMessageContent1.CreateStr(
                "Answer as a technical writer. Keep the reply concise."),
        }),
        Item.CreateInputMessage(new InputMessage()
        {
            Role = InputMessageRole.User,
            Content = InputMessageContent1.CreateStr(
                "Draft a short product update about faster response times."),
        }),
    }),
});

mka1 llm responses create --body '{
  "model": "auto",
  "input": [
    { "type": "message", "role": "developer", "content": "Answer as a technical writer. Keep the reply concise." },
    { "type": "message", "role": "user", "content": "Draft a short product update about faster response times." }
  ]
}'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "input": [
      { "type": "message", "role": "developer", "content": "Answer as a technical writer. Keep the reply concise." },
      { "type": "message", "role": "user", "content": "Draft a short product update about faster response times." }
    ]
  }'

This pattern is useful when you want the request body to carry the message history directly.

Continue a multi-turn exchange

Use previous_response_id to continue from an earlier response without resending the full history.

const second = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    previousResponseId: 'resp_123',
    input: 'Now turn that into an email subject line.',
  },
});

const second = await openai.responses.create({
  model: 'auto',
  previous_response_id: 'resp_123',
  input: 'Now turn that into an email subject line.',
  stream: false,
});

res = sdk.llm.responses.create(
    model="auto",
    previous_response_id="resp_123",
    input="Now turn that into an email subject line.",
)

// First request
var first = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Input = ResponsesCreateRequestInput.CreateStr("Write a one-line product tagline."),
});

// Second request: continue from the first response
var second = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    PreviousResponseId = first.Id,
    Input = ResponsesCreateRequestInput.CreateStr("Now turn that into an email subject line."),
});

mka1 llm responses create \
  --model auto \
  --previous-response-id resp_123 \
  --input '"Now turn that into an email subject line."'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "previous_response_id": "resp_123",
    "input": "Now turn that into an email subject line."
  }'

If you need a reusable conversation container, create one with the Conversations resource and then pass the conversation ID in conversation.

const conv = await sdk.llm.conversations.create({
  createConversationRequest: {
    metadata: { session_id: 'web-42' },
  },
});

const result = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    conversation: conv.id,
    input: 'What should I ask next to refine this draft?',
  },
});

const conv = await openai.conversations.create({
  metadata: { session_id: 'web-42' },
});

const response = await openai.responses.create({
  model: 'auto',
  conversation: conv.id,
  input: 'What should I ask next to refine this draft?',
  stream: false,
});

conv = sdk.llm.conversations.create(
    create_conversation_request={
        "metadata": {"session_id": "web-42"},
    }
)

res = sdk.llm.responses.create(
    model="auto",
    conversation=conv.id,
    input="What should I ask next to refine this draft?",
)

var conv = await sdk.Llm.Conversations.CreateAsync(body: new CreateConversationRequest()
{
    Metadata = new Dictionary<string, string> { { "session_id", "web-42" } },
});

# Create a conversation
mka1 llm conversations create --body '{
  "metadata": { "session_id": "web-42" }
}'

# Use the conversation in a response request
mka1 llm responses create \
  --model auto \
  --conversation conv_123 \
  --input '"What should I ask next to refine this draft?"'

# Create conversation
curl https://apigw.mka1.com/api/v1/llm/conversations \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "metadata": { "session_id": "web-42" }
  }'

# Use conversation in a response request
curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "conversation": "conv_123",
    "input": "What should I ask next to refine this draft?"
  }'

Note: deleting a conversation does not delete its items. Deleting a response permanently deletes the response and associated data. See the Conversations and Responses pages in the API Reference for the full resource workflow.

Stream a response

See Streaming for complete client examples.

Tune sampling and length

Sampling fields change how the model picks each token. Length fields cap how much it writes. All of them are optional and sit at the top level of the request, next to model and input.

const result = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    input: 'Write a one-sentence summary of the MKA1 API.',
    temperature: 0.2,
    maxOutputTokens: 400,
  },
});

const response = await openai.responses.create({
  model: 'auto',
  input: 'Write a one-sentence summary of the MKA1 API.',
  temperature: 0.2,
  max_output_tokens: 400,
  stream: false,
});

res = sdk.llm.responses.create(
    model="auto",
    input="Write a one-sentence summary of the MKA1 API.",
    temperature=0.2,
    max_output_tokens=400,
)

var res = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Input = ResponsesCreateRequestInput.CreateStr(
        "Write a one-sentence summary of the MKA1 API."),
    Temperature = 0.2,
    MaxOutputTokens = 400,
});

mka1 llm responses create --body '{
  "model": "auto",
  "input": "Write a one-sentence summary of the MKA1 API.",
  "temperature": 0.2,
  "max_output_tokens": 400
}'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "input": "Write a one-sentence summary of the MKA1 API.",
    "temperature": 0.2,
    "max_output_tokens": 400
  }'

The gateway validates every value before it calls the model. A value outside its range returns 400. In the example, check status: if the model hit the 400-token cap, the response is incomplete rather than completed, and usage.output_tokens stays at or below the cap. Some fields below are gateway extensions. They are not part of the OpenAI Responses API, and whether they reach the model depends on its upstream, the provider API the gateway forwards your request to, which speaks either the Responses format or the Completions format.

Field	Accepts	Effect
`temperature`	number, 0 to 2	How random each token pick is. Lower values give steadier output.
`top_p`	number, 0 to 1	Nucleus sampling. The model only considers the tokens that make up the top `top_p` share of probability. Use this or `temperature`, not both.
`max_output_tokens`	integer, 1 or more	Cap on generated tokens. A response cut off at the cap finishes with `status: "incomplete"`.
`stop`	string, or a non-empty array of strings	Stop sequences. Gateway extension.
`presence_penalty`	number, -2 to 2	Penalizes tokens that have already appeared, so the model moves on to new topics.
`frequency_penalty`	number, -2 to 2	Penalizes tokens by how often they have appeared, so the model repeats itself less.
`top_k`	integer, 0 or more	Provider-specific top-k sampling. Gateway extension.
`min_p`	number, 0 to 1	Provider-specific min-p sampling. Gateway extension.
`repetition_penalty`	number, 0 or more	Provider-specific repetition penalty. Gateway extension.
`do_sample`	boolean	Provider-specific sampling toggle. Gateway extension.
`truncation`	`"auto"` or `"disabled"`, default `"auto"`	What happens when the input outgrows the context window. `auto` drops the oldest messages; with `disabled` the response fails with status `failed`.

Fields marked as gateway extensions are not part of the OpenAI Responses API, so the OpenAI SDK’s types do not know them. Send them with the MKA1 SDK, the CLI, or raw HTTP. Four more gateway extensions pass provider options through untouched: chat_template_kwargs (for example {"enable_thinking": false}), prefill_think (a boolean or prefill text), use_cache, and extra_body, an object merged into the upstream request after the gateway strips any auth, base-URL, and header keys from it. Not every field reaches every model. Fields the gateway drops lists the cases.

Check what a model supports

Not every model accepts every field. GET /api/v1/llm/models/{model_id} returns a capabilities object with four booleans:

Flag	`true` means
`supports_temperature`	The model accepts `temperature`.
`supports_top_p`	The model accepts `top_p`.
`supports_reasoning`	The model reasons before it answers, so part of its output tokens go to reasoning.
`supports_reasoning_effort`	The model accepts `reasoning.effort`. A reasoning model can report `false` here: it still reasons, but ignores the effort level.

Models that are not LLMs (embeddings, images, speech) report false for all four. The model object also lists input_modalities and output_modalities, the kinds of content the model accepts and produces, such as ["text", "image"].

const model = await sdk.llm.models.get({ modelId: 'meetkai:functionary' });

console.log(model.capabilities, model.inputModalities);

const model = await openai.models.retrieve('meetkai:functionary');

// capabilities and input_modalities are gateway extensions, so the OpenAI SDK's Model type does not declare them
const { capabilities, input_modalities } = model as unknown as {
  capabilities: {
    supports_temperature: boolean;
    supports_top_p: boolean;
    supports_reasoning: boolean;
    supports_reasoning_effort: boolean;
  };
  input_modalities: string[];
};
console.log(capabilities, input_modalities);

model = sdk.llm.models.get(model_id="meetkai:functionary")

print(model.capabilities, model.input_modalities)

var model = await sdk.Llm.Models.GetAsync(modelId: "meetkai:functionary");

var capabilities = model.Object!.Capabilities;
Console.WriteLine($"{capabilities.SupportsTemperature} {capabilities.SupportsTopP} {capabilities.SupportsReasoning} {capabilities.SupportsReasoningEffort}");
Console.WriteLine(string.Join(", ", model.Object!.InputModalities));

mka1 llm models get --model-id meetkai:functionary

curl https://apigw.mka1.com/api/v1/llm/models/meetkai:functionary \
  --header 'Authorization: Bearer <mka1-api-key>'

If you send a field to a model that reports false for it, the gateway does not reject the request. It drops that one field before calling the model, forwards everything else, and the model uses its own default for that setting. Nothing in the response says the field was dropped: it echoes temperature, top_p, and reasoning exactly as you sent them. Check the model’s capabilities first when the setting matters.

Control reasoning

Reasoning models think before they answer. The reasoning object sets how much thinking you want and whether you see it:

effort: none, minimal, low, medium, high, or xhigh.
summary: auto, concise, or detailed. Asks the model to return a readable summary of its reasoning.

Both keys must be present whenever you send the object. Set the one you are not using to null. {"effort": "low"} on its own fails validation. Each key reaches the model only when the model can use it:

effort goes only to models that report supports_reasoning_effort: true. A Completions-format upstream receives it as reasoning_effort.
summary goes only to models that report supports_reasoning: true and whose upstream speaks the Responses format.

The gateway drops a key the model cannot use, and the request still succeeds.

const result = await sdk.llm.responses.create({
  responsesCreateRequest: {
    model: 'auto',
    input: 'Which of 17, 21, and 23 are prime? Answer in one line.',
    reasoning: { effort: 'low', summary: null },
  },
}) as ResponseObject;

console.log(result.usage.outputTokensDetails.reasoningTokens);

const response = await openai.responses.create({
  model: 'auto',
  input: 'Which of 17, 21, and 23 are prime? Answer in one line.',
  reasoning: { effort: 'low', summary: null },
  stream: false,
});

console.log(response.usage?.output_tokens_details.reasoning_tokens);

res = sdk.llm.responses.create(
    model="auto",
    input="Which of 17, 21, and 23 are prime? Answer in one line.",
    reasoning={"effort": "low", "summary": None},
)

print(res.usage.output_tokens_details.reasoning_tokens)

var res = await sdk.Llm.Responses.CreateAsync(new ResponsesCreateRequest()
{
    Model = "auto",
    Input = ResponsesCreateRequestInput.CreateStr(
        "Which of 17, 21, and 23 are prime? Answer in one line."),
    Reasoning = new ReasoningConfiguration()
    {
        Effort = ReasoningConfigurationEffort.Low,
        Summary = null,
    },
});

Console.WriteLine(res.ResponseObject!.Usage.OutputTokensDetails.ReasoningTokens);

mka1 llm responses create --body '{
  "model": "auto",
  "input": "Which of 17, 21, and 23 are prime? Answer in one line.",
  "reasoning": { "effort": "low", "summary": null }
}'

curl https://apigw.mka1.com/api/v1/llm/responses \
  --request POST \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <mka1-api-key>' \
  --header 'X-On-Behalf-Of: <end-user-id>' \
  --data '{
    "model": "auto",
    "input": "Which of 17, 21, and 23 are prime? Answer in one line.",
    "reasoning": { "effort": "low", "summary": null }
  }'

The OpenAI SDK’s ReasoningEffort type does not include none or xhigh; send those with the MKA1 SDK, the CLI, or raw HTTP. Reasoning shows up in three places:

output can contain an item with type: "reasoning". Its summary array holds summary_text parts, and content (when present) holds the raw reasoning_text parts. Add "reasoning.encrypted_content" to include to get encrypted_content as well.
usage.output_tokens_details.reasoning_tokens is the share of output_tokens the model spent thinking. Reasoning tokens are output tokens, so they count against max_output_tokens.
In a stream, response.reasoning_text.delta and response.reasoning_summary_text.delta events typically arrive before the response.output_text.delta events. The streaming output latency benchmark shows this order on a live model.

To know before the call whether a model reasons and whether it accepts effort, read supports_reasoning and supports_reasoning_effort as shown in Check what a model supports. After the call, usage.output_tokens_details.reasoning_tokens shows how much reasoning happened. When you send model: "auto", the gateway applies the organization’s auto policy for Responses. The default is a direct model-registry alias. An organization admin can configure a balanced policy that asks an LLM judge to choose among eligible models using capability, price, gateway latency, comparable evals, request requirements, and conversation continuity. Routing does not change the reasoning.effort you send, but the gateway still drops it when the chosen model reports supports_reasoning_effort: false. Auto routing explains the policy and its failure and privacy behavior.

Fields the gateway drops

Each field below passes validation, but the gateway drops it before calling the model in the case listed. A dropped field never fails the request, and nothing in the response says it was dropped. Some cases depend on the model’s upstream, the provider API the gateway forwards your request to.

Field	Dropped when
`temperature`	The model reports `supports_temperature: false`.
`top_p`	The model reports `supports_top_p: false`.
`reasoning.effort`	The model reports `supports_reasoning_effort: false`.
`reasoning.summary`	The model reports `supports_reasoning: false`, or its upstream speaks the Completions format.
`text.verbosity`	The model’s upstream speaks the Completions format.
`stop`	The model’s upstream speaks the Responses format at `api.openai.com`.
`top_k`, `min_p`, `repetition_penalty`, `do_sample`, `chat_template_kwargs`, `prefill_think`, `use_cache`, `extra_body`	The model’s upstream is at `api.openai.com`.
`top_logprobs`, `service_tier`, `prompt_cache_key`, `safety_identifier`, `user`	Always. The gateway accepts them for OpenAI compatibility but does not send them to the model.
Any field the Responses API does not define	Always. The gateway removes it without an error.

Because unknown fields are removed silently, a misspelled field name looks accepted but has no effect. If a setting seems to do nothing, compare your request with the API reference.

Pause a background response

A background response can call the sleep tool to wait up to 20 minutes for a webhook or an outside job, and you can wake it early with follow-up input. See background responses.

Next steps

Review the API overview for authentication and base URL details
See background responses when you need to offload long-running work and poll or stream for results
See manage conversations to organize multi-turn exchanges into reusable conversation containers
See manage agents when you want reusable agent definitions and persisted runs
See streaming output latency for measured time to first streamed token on the Responses API

Get started

SDKs and CLI

Responses

Standalone APIs

Agents

Evals

Compute

Authentication

Production

Text generation

Send a simple prompt

Add instructions

Send structured messages

Continue a multi-turn exchange

Stream a response

Tune sampling and length

Check what a model supports

Control reasoning

Fields the gateway drops

Pause a background response

Next steps

​Send a simple prompt

​Add instructions

​Send structured messages

​Continue a multi-turn exchange

​Stream a response

​Tune sampling and length

​Check what a model supports

​Control reasoning

​Fields the gateway drops

​Pause a background response

​Next steps

Send a simple prompt

Add instructions

Send structured messages

Continue a multi-turn exchange

Stream a response

Tune sampling and length

Check what a model supports

Control reasoning

Fields the gateway drops

Pause a background response

Next steps