openrouter(options)

Creates an OpenRouterProvider that implements AIProvider using the OpenRouter API — a unified endpoint for 200+ models.

Signature

function openrouter(options: OpenRouterOptions): OpenRouterProvider

interface OpenRouterOptions {
  /** OpenRouter API key (sk-or-...) */
  apiKey: string;
  /** Default model identifier. Optional — overridden by per-agent model. */
  defaultModel?: string;
  /** API base URL. Defaults to 'https://openrouter.ai/api/v1'. */
  baseUrl?: string;
  /** Extra headers sent with every request. */
  headers?: Record<string, string>;
}

Usage

import { openrouter, configure } from '@daedalus-ai-dev/ai-sdk';

configure({
  provider: openrouter({
    apiKey: process.env.OPENROUTER_API_KEY!,
    headers: {
      'HTTP-Referer': 'https://myapp.com',  // Recommended for OpenRouter analytics
      'X-Title': 'My Application',
    },
  }),
  model: 'openai/gpt-4o-mini',
});

Selecting models

OpenRouter uses the format provider/model-name. Common choices:

Model	Identifier	Best for
GPT-4o mini	openai/gpt-4o-mini	Fast tasks, classification, simple Q&A
GPT-4o	openai/gpt-4o	Balanced capability and speed
Claude 3.5 Sonnet	anthropic/claude-3-5-sonnet	Complex reasoning, long-form writing
Claude 3 Haiku	anthropic/claude-3-haiku	Ultra-fast, cheap
Gemini Flash 1.5	google/gemini-flash-1.5	Large context, fast
Llama 3.3 70B	meta-llama/llama-3.3-70b-instruct	Open-source, cost-effective

See the full list at openrouter.ai/models.

Structured output (JSON Schema)

When a schema is provided and no tools are active, the provider sends the schema as a response_format with strict: true. This constrains the model to return valid JSON matching the schema.

const response = await agent({
  instructions: 'Extract product details.',
  schema: (s) => ({
    name:  s.string().required(),
    price: s.number().min(0).required(),
  }),
}).prompt('Blue Widget — $12.99');

// response.structured is parsed JSON

Model compatibility

Not all models support response_format: json_schema. GPT-4o and GPT-4o-mini have the best structured output support. For other models, consider asking the model to return JSON in the prompt and parsing response.text manually.

Tool calling

The provider maps Tool[] to OpenAI-compatible function definitions and handles tool call parsing automatically. Tool calls within a single turn are executed in parallel.

Streaming

The provider implements Server-Sent Events (SSE) streaming and reassembles tool call deltas across chunks transparently.

Using a self-hosted endpoint

Point to any OpenAI-compatible API:

const provider = openrouter({
  apiKey: 'my-local-key',
  baseUrl: 'http://localhost:11434/v1',  // e.g. Ollama
});

OpenRouterProvider class

You can also instantiate the class directly if you prefer:

import { OpenRouterProvider } from '@daedalus-ai-dev/ai-sdk';

const provider = new OpenRouterProvider({
  apiKey: process.env.OPENROUTER_API_KEY!,
});