OrcaRouter provider for the Vercel AI SDK

OrcaRouter is an OpenAI-compatible meta-router: one API key reaches 150+ models across OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, Kimi, MiniMax, Z-AI and others. It also exposes orcarouter/auto, an adaptive router that picks an upstream per request based on your configured strategy (cheapest, balanced, quality, adaptive, gated_adaptive).

This package is the OrcaRouter provider for the Vercel AI SDK. It wraps @ai-sdk/openai-compatible and reshapes outgoing requests so per-vendor reasoning protocols and upstream parameter quirks are handled for you.

Installation

npm install orcarouter-ai-sdk-provider

Usage

import { createOrcaRouter } from "orcarouter-ai-sdk-provider"
import { generateText } from "ai"

const orcarouter = createOrcaRouter({
	apiKey: process.env.ORCAROUTER_API_KEY, // get one at https://www.orcarouter.ai/console/keys
})

const { text } = await generateText({
	model: orcarouter("orcarouter/auto"),
	prompt: "Explain adaptive routing in one sentence.",
})

Or use the default instance, which reads ORCAROUTER_API_KEY from the environment:

import { orcarouter } from "orcarouter-ai-sdk-provider"

const model = orcarouter("anthropic/claude-opus-4.7")

Configuration

createOrcaRouter(settings) accepts:

Option	Description
apiKey	OrcaRouter API key. Defaults to process.env.ORCAROUTER_API_KEY.
baseURL	API base URL. Defaults to https://api.orcarouter.ai/v1.
headers	Extra headers sent with every request.
appUrl / appName	Optional attribution shown in the OrcaRouter console (sets HTTP-Referer / X-Title).
fetch	Custom fetch implementation (e.g. for proxies or tests).

Models

Pick any model by its OrcaRouter id, for example:

• orcarouter/auto — adaptive routing (recommended default)
• anthropic/claude-fable-5, anthropic/claude-opus-4.8, anthropic/claude-sonnet-4.6
• openai/gpt-5.5, openai/gpt-5-mini
• google/gemini-3.5-flash
• grok/grok-4.3, deepseek/deepseek-v4-pro, minimax/minimax-m2.7
• deepseek/deepseek-reasoner
• kimi/kimi-k2.6, z-ai/glm-4.5, qwen/qwen3.7-max

The full catalog and live pricing are at orcarouter.ai/models.

Reasoning

Pass a reasoning effort via the AI SDK and this provider forwards it in each upstream's native shape:

• Anthropic Claude thinking models receive a top-level thinking block (the effort is mapped to a token budget).
• OpenAI / Gemini / Grok / Qwen / Kimi receive a flat reasoning_effort.
• DeepSeek r1 / reasoner reason automatically; reasoning control fields are stripped.

Upstream parameter quirks are also handled automatically — for example temperature is omitted for orcarouter/* routers, Claude Opus 4.7 / 4.8, Claude Fable 5, and the gpt-5 family (which reject it), and Kimi K2.6 is pinned to its only accepted temperature / top_p.

Notes on orcarouter/auto

The adaptive router's candidate pool may include upstreams that don't support function/tool calling. If a tool-calling request fails with a "function calling not enabled" error, pin a tool-capable model (e.g. anthropic/claude-opus-4.8) or adjust the pool in the routing console.

License

Apache-2.0