@remcostoeten/ai-core owns the contracts and the run lifecycle. This package owns what it takes to reach a real provider, and its whole job is that the vendor library stays an implementation detail: no LanguageModel, UIMessage, TextStreamPart or APICallError appears in anything it exports. Vendor packages are optional peer dependencies, so you install only the one you use.

• Named factories only: createGroqProvider, createGeminiProvider, createAnthropicProvider, createOpenAiCompatibleProvider. There is no fromLanguageModel(unknown) escape hatch
• Ordering is a guarantee. The model authority is consulted, then the credential resolves, then a socket opens, so an unauthorized model or an unconfigured provider terminalizes with nothing sent
• maxRetries: 0, always. A 429 is attempted exactly once, because a hidden retry spends budget and time the caller allotted to one attempt
• No provider response body crosses the boundary. Bodies echo prompts and, from some vendors, a key prefix; only a bounded, status-derived sentence reaches the caller
• Credentials arrive through a CredentialSource port and model permission through a ModelAuthority port, both yours to implement
• Nothing here reads process.env or Deno.env, so the package makes no assumption about which runtime it is in

Each of those is asserted by a test, including the one that greps the sources for environment access and the one that asserts the fixture fetch was never called on a refusal.

The adapter is documented in typescript-core.md, worked usage is in examples.md, and the reasoning behind the boundary is ADR 0003.

Install

Install this package, ai-core, and the vendor package for the provider you reach.

npm install @remcostoeten/ai-sdk @remcostoeten/ai-core @ai-sdk/groq

The optional peers are @ai-sdk/groq, @ai-sdk/google, @ai-sdk/anthropic and @ai-sdk/openai-compatible.

Use

import { buildRequest, consumeEvents, createRuntime } from '@remcostoeten/ai-core'
import { createGroqProvider, staticCredential } from '@remcostoeten/ai-sdk'

const provider = await createGroqProvider({
  credentials: staticCredential(process.env.GROQ_API_KEY ?? ''),
  models: { permits: (_, modelId) => modelId === 'llama-3.3-70b-versatile' },
})

const runtime = createRuntime({ providers: [provider] })

From there the run is @remcostoeten/ai-core: build a request, stream it, and consume the events.

Not implemented

No retries and no key rotation. No fallback or routing between providers. No structured output. No model listing or capability discovery. Every adapter test runs against a fixture fetch, so nothing here has been confirmed against a live endpoint.

xxx, Remco Stoeten MIT