Supercompat is an AI compatibility layer without compromises. A TypeScript library that lets you call any LLM provider through the OpenAI SDK (or the Anthropic SDK). Swap one adapter and the same client.responses.create() call reaches Anthropic, Google, Groq, Mistral, Together, OpenRouter, Perplexity, Ollama, or Azure — with the original SDK types intact.

It runs in-process. No proxy server, no request forwarding, no extra latency. Supercompat installs a custom fetch on the SDK instance and routes calls locally.

Full docs: supercompat.com/docs.

Install

npm install supercompat openai

Quick example

import {
  supercompat,
  anthropicClientAdapter,
  completionsRunAdapter,
  memoryStorageAdapter,
} from 'supercompat/openai'
import Anthropic from '@anthropic-ai/sdk'

const client = supercompat({
  clientAdapter: anthropicClientAdapter({ anthropic: new Anthropic() }),
  storageAdapter: memoryStorageAdapter(),
  runAdapter: completionsRunAdapter(),
})

const response = await client.responses.create({
  model: 'claude-sonnet-4-6',
  input: 'Say hello.',
})

console.log(response.output_text)

client is a real OpenAI instance with the real TypeScript types. Every call made on it — responses, chat.completions, beta.threads — is intercepted by Supercompat and translated into a request against the Anthropic SDK. Switching providers is a change to clientAdapter; everything else stays the same.

Persistent state

memoryStorageAdapter is fine for one-shot scripts but loses everything on restart. For persisted conversations, threads, and runs, swap it for prismaStorageAdapter:

import { PrismaClient } from '@prisma/client'
import {
  supercompat,
  anthropicClientAdapter,
  completionsRunAdapter,
  prismaStorageAdapter,
} from 'supercompat/openai'
import Anthropic from '@anthropic-ai/sdk'

const prisma = new PrismaClient()

const client = supercompat({
  clientAdapter: anthropicClientAdapter({ anthropic: new Anthropic() }),
  storageAdapter: prismaStorageAdapter({ prisma }),
  runAdapter: completionsRunAdapter(),
})

// Continue a conversation across requests with previous_response_id:
const first = await client.responses.create({
  model: 'claude-sonnet-4-6',
  input: 'My name is Alice.',
})

const second = await client.responses.create({
  model: 'claude-sonnet-4-6',
  input: 'What did I just tell you?',
  previous_response_id: first.id,
})

Conversations, responses, assistants, threads, messages, and runs all land in Postgres. See Storage adapters for every option — including OpenAI-managed and Azure-managed state.

Where to go next

• Installation — install the package, pick a provider SDK, and wire them together.
• Comparison — how Supercompat compares to Vercel AI SDK, LiteLLM, LangChain, and others.
• Output SDKs — return an OpenAI-shaped or Anthropic-shaped client. Works with every provider.
• Adapters — the three adapter types (client, storage, run) and how they compose.
• Providers — setup notes for OpenAI, Anthropic, Google, Azure, and every other backend.
• Tools — function calling, web search, file search, code interpreter, and computer use.
• Streaming — stream deltas through the OpenAI SDK regardless of which provider is behind it.

Monorepo layout

This repository is a Turborepo workspace. The published npm package lives at packages/supercompat — that's the one most consumers care about. The docs/ directory holds the MDX source that renders on supercompat.com/docs.

Links

• Docs: supercompat.com/docs
• npm: npmjs.com/package/supercompat
• GitHub: github.com/supercorp-ai/supercompat
• Supported by Supercorp