OpenRouter AI SDK Provider

OpenRouter AI SDK Provider is the official adapter that plugs OpenRouter into the Vercel AI SDK, giving your application a single, type-safe entry point to more than 300 large language models from OpenAI, Anthropic, Google, Meta, and dozens of other providers. Instead of wiring up a separate SDK for each vendor, you import one openrouter instance and pass a model id like openai/gpt-4o or anthropic/claude-sonnet-4 to the AI SDK’s generateText, streamText, generateObject, and embed functions.

Because it implements the AI SDK provider interface, it drops into any existing AI SDK codebase without new call patterns, while exposing OpenRouter-specific capabilities such as usage accounting, prompt caching, provider routing, response healing, and per-request extra body. It is ESM-only, ships full TypeScript types, and tracks the AI SDK v7 release line.

What You Get

• A drop-in Vercel AI SDK provider that exposes 300+ models through one import
• Full TypeScript type definitions and ESM-only, tree-shakeable output
• Chat, completion, embedding, image, and video model factories under one client
• OpenRouter-native features: usage accounting, prompt caching, provider routing, and response healing
• Per-request and per-model extra body plus custom header passthrough for provider-specific options

Common Use Cases

• Switching between LLM vendors without rewriting application code
• Building streaming chat and agentic tool-use features on top of the AI SDK
• Generating structured JSON output with schema validation and response healing
• Powering RAG and semantic search with OpenRouter embedding models

Under The Hood

Architecture — The package is organized by modality: src/facade.ts exposes a deprecated OpenRouter class while src/provider.ts builds the modern createOpenRouter factory that returns a callable OpenRouterProvider (invoking it directly returns a chat model). Each modality lives in its own directory (chat, completion, embedding, image, video), and every factory constructs a language-model class (e.g. OpenRouterChatLanguageModel) with a shared base config that lazily assembles auth and app-identity headers via loadApiKey, injecting X-OpenRouter-Title, HTTP-Referer, and optional X-Provider-API-Keys. src/index.ts re-exports the facade, provider, and types.

Tech Stack — Written in TypeScript (99% of the codebase), it declares peer dependencies on ai@^7.0.0 and zod, and builds against @ai-sdk/provider and @ai-sdk/provider-utils. It is bundled with tsup, linted and formatted with Biome, tested with Vitest across node, edge, e2e, and issues configs, and released via Changesets with npm provenance. Output is ESM-only, side-effect-free, and requires Node.js 22+.

Code Quality — The repo carries roughly 21 in-tree unit tests plus around 56 end-to-end tests, exercising node and edge runtimes with MSW-based request mocking. Public options are documented with JSDoc, types are strict, and deprecated surfaces (the OpenRouter class, embedding()) are clearly annotated with forward pointers to their replacements.

API Design — The developer experience is deliberately minimal: import { openrouter } and call openrouter('openai/gpt-4o'), or createOpenRouter({ apiKey }) for configuration. Model ids are typed, provider-specific behavior flows through providerOptions.openrouter, extraBody, and headers at either the factory or request level, and because it conforms to the AI SDK provider contract, existing generateText/streamText/generateObject/embed code works with zero new call patterns.