runtime.response()

One-shot prompt-in / JSON-out call with MCP servers, reasoning effort, and structured output

runtime.response() is an ergonomic wrapper for the common "give me one answer back" case — a single prompt, optional MCP servers, optional reasoning effort, optional JSON schema, one response. No streaming, no multi-step agent loop.

It's a thin convenience over runtime.generate(): same adapter chain, same per-provider translation, same fallback behavior. Use it when you don't need streaming or tool-call introspection.

import { createRuntime, createFallbackChain } from '@yourgpt/llm-sdk';
import { createOpenAI } from '@yourgpt/llm-sdk/openai';
import { createAnthropic } from '@yourgpt/llm-sdk/anthropic';

const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY });
const anthropic = createAnthropic({ apiKey: process.env.ANTHROPIC_API_KEY });

const runtime = createRuntime({
  adapter: createFallbackChain({
    models: [
      openai.languageModel('gpt-4o'),
      anthropic.languageModel('claude-opus-4-7'),
    ],
    strategy: 'priority',
  }),
});

const result = await runtime.response({
  prompt: 'Extract the top 3 fastest land animals as JSON.',
  reasoningEffort: 'medium',
  responseFormat: {
    type: 'json_schema',
    json_schema: {
      name: 'animals',
      schema: {
        type: 'object',
        properties: {
          animals: {
            type: 'array',
            items: {
              type: 'object',
              properties: {
                name: { type: 'string' },
                top_speed_kmh: { type: 'number' },
              },
              required: ['name', 'top_speed_kmh'],
            },
          },
        },
        required: ['animals'],
      },
      strict: true,
    },
  },
});

const data = JSON.parse(result.text);
// → { animals: [{ name: 'Cheetah', top_speed_kmh: 120 }, ...] }

Parameters

Param	Type	Description
`prompt`	`string`	The user message. Sent as a single `role: 'user'` turn.
`systemPrompt`	`string?`	Optional system prompt (overrides any default on the runtime).
`mcpServers`	`McpServerConfig[]?`	Remote MCP servers the model can call. See Structured Output → MCP servers.
`reasoningEffort`	`ReasoningEffort?`	`'minimal'`/`'low'`/`'medium'`/`'high'` or `{ budgetTokens }` or `{ raw }`. See Structured Output → Reasoning effort.
`responseFormat`	`ResponseFormat?`	JSON-schema constraint. See Structured Output.
`maxTokens`	`number?`	Maximum tokens for the completion.
`temperature`	`number?`	Sampling temperature (0–2). Silently dropped on OpenAI reasoning models.

Result

{
  text: string;        // model output (parse yourself if responseFormat used)
  toolCalls: Array<{   // function-tool calls; MCP tool calls are resolved provider-side
    id: string;
    name: string;
    args: Record<string, unknown>;
  }>;
}

The method throws if the underlying call errors out. Catch at the call site or rely on the fallback chain to route past failures.

When to use `response()` vs `generate()` / `chat()`

Use case	Method
One-shot prompt, want JSON back	`runtime.response()`
Multi-turn conversation, full message history	`runtime.generate()`
Streaming output to the client	`runtime.stream()` / `streamText()`
Custom tool definitions with handlers, agent loop	`runtime.chat()` with `tools`

If your call only needs prompt + responseFormat + maxTokens, response() reads cleaner. The other methods give you more knobs for free.

Provider routing under the hood

response() forwards through the standard adapter pipeline:

OpenAI — when mcpServers or reasoningEffort is set, routes through /v1/responses (Chat Completions doesn't accept these fields). responseFormat becomes text.format.
Anthropic — uses client.beta.messages.* with the mcp-client-2025-11-20 beta header when MCP servers are present. reasoningEffort maps to adaptive thinking on Claude 4.6/4.7 and budget_tokens on older models.
Google / xAI / OpenRouter — these share the OpenAI adapter via an OpenAI-compatible endpoint. They throw a clear error when mcpServers or reasoningEffort is set, so fallback chains skip past them.

The store: false flag is always sent on /v1/responses calls — response() is one-shot semantics by design, no conversation persistence on OpenAI's side. If you need persistence, use runtime.chat() instead.

Self-learning example

The canonical use case: extract FAQs from a chat transcript, consulting a knowledge-base MCP server first to avoid duplicates.

const FAQ_SCHEMA = {
  type: 'object',
  properties: {
    response: {
      type: 'array',
      items: {
        type: 'object',
        properties: {
          question: { type: 'string' },
          answer: { type: 'string' },
          intent: { type: 'string' },
        },
        required: ['question', 'answer', 'intent'],
        additionalProperties: false,
      },
    },
  },
  required: ['response'],
  additionalProperties: false,
} as const;

const result = await runtime.response({
  prompt,
  systemPrompt: 'Extract FAQs. Consult the KB before creating entries.',
  mcpServers: [{
    label: 'knowledge_base',
    url: 'https://kb.example.com/sse',
    headers: { Authorization: `Bearer ${env.KB_TOKEN}` },
    allowedTools: ['search_kb'],
    requireApproval: 'never',
  }],
  reasoningEffort: 'high',
  responseFormat: {
    type: 'json_schema',
    json_schema: { name: 'faqs', schema: FAQ_SCHEMA, strict: true },
  },
});

const parsed = JSON.parse(result.text).response;

A working version lives in examples/fallback-demo — see the POST /response route.

Next Steps

Structured Output — full reference for responseFormat, mcpServers, reasoningEffort
generateText() — multi-turn / tool-loop API
Tools — function-calling primitives