@openagentry/adapter-llm-aigateway

LlmAdapter implementation backed by Vercel AI Gateway via the official ai SDK.

Pairs naturally with @openagentry/adapter-host-vercel and @openagentry/adapter-storage-vercel-blob for projects deployed on Vercel; can be used standalone with any gateway-bearing client.

Install

pnpm add @openagentry/adapter-llm-aigateway

Quick start

import llm from '@openagentry/adapter-llm-aigateway';

const models = await llm.listModels();
const r = await llm.generate('Say hi!', { model: 'anthropic/claude-opus-4.6' });
console.log(r.text);

The default export reads OA_AIGATEWAY_API_KEY from process.env lazily on first call. Or construct an adapter explicitly:

import { createAigatewayLlmAdapter } from '@openagentry/adapter-llm-aigateway';

const llm = createAigatewayLlmAdapter({
  apiKey: process.env.MY_KEY!,
  defaultModel: 'anthropic/claude-opus-4.6',
});

Configuration

| Env var | Required | Default | Notes | |---|---|---|---| | OA_AIGATEWAY_API_KEY | yes | — | Vercel AI Gateway key (Vercel dashboard → AI Gateway → Keys). | | OA_AIGATEWAY_DEFAULT_MODEL | no | — | Model id ('provider/model'). No built-in default. | | OA_AIGATEWAY_BASE_URL | no | https://ai-gateway.vercel.sh/v1 | Gateway base URL. |

We use OA_AIGATEWAY_API_KEY (not Vercel's own AI_GATEWAY_API_KEY) to keep adapter env naming consistent with the rest of OpenAgentry. The adapter sets AI_GATEWAY_API_KEY from your OA_AIGATEWAY_API_KEY automatically before calling the SDK, so the SDK reads the right value.

Operator vs runtime flavors

• Operator flavor (@openagentry/adapter-llm-cli) — shells out to local claude/codex/gemini CLIs. No API key. Build-time only.
• Runtime flavor (this package) — programmatic gateway access. Requires API key. Works in deployed functions.

Both implement the same LlmAdapter interface; pick the one your runtime needs.

What's NOT in v0

• Tool / function calling
• Vision / multimodal inputs
• Embeddings (separate EmbeddingsAdapter in future)
• system as message-array (string only)
• Provider-specific options (Anthropic prompt caching, OpenAI logit_bias, etc.)
• Web search / observability / cost tracking

For these features today, drop down to the raw ai SDK directly.

LLM-facing docs

See LLM.md — the doc shipped to LLM consumers via package.json#openagentry.llmDocs.

License

Apache-2.0