Mainstream providers through AI SDK

Since @lite-agent/provider 0.11.0, aiSdk(model, options?) bridges AI SDK 7 LanguageModelV4 models. It requires Node.js >=22 and rejects legacy V2/V3 contracts.

The bridge normalizes requests and streams; maintained vendor adapters own HTTP, authentication and wire formats. Vendor packages below are installed by the host as needed, not bundled as production dependencies. There is no second agent loop, implicit Gateway, extra retry layer or automatic input-URL download.

Quick start

pnpm add @lite-agent/sdk @lite-agent/provider @ai-sdk/google

import { createLiteAgent } from '@lite-agent/sdk';
import { aiSdk } from '@lite-agent/provider';
import { createGoogle } from '@ai-sdk/google';

const google = createGoogle({ apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY });
const model = google(process.env.GEMINI_MODEL!);
const agent = createLiteAgent({
  workdir: process.cwd(),
  model: aiSdk(model, { context: { contextWindow: 128_000 } }),
});
try {
  console.log((await agent.send('Analyze this project')).text);
} finally {
  await agent.close();
}

Set the window to the actual model limit; it is not inferred from model names. The bridge provider id is its bound modelId, so ordinary calls can omit modelName. Explicit names must match. For multi-agent models profiles, bind a matching provider instance for each model instead of silently ignoring routing choices.

Coverage

The SDK supported providers page owns the vendor matrix, packages, factories, compatible endpoints and verification scope. See supported protocols for protocol distinctions. This page covers bridge configuration and replay behavior.

Behavior

Maps token limits, temperature, topP, stop sequences, seed, function schemas and tool choice to standard V4 parameters.
Maps reasoningEffort to V4 reasoning; vendor adapters handle their fields. Main and child agents can use different efforts. Dedicated tests cover high effort for OpenAI, Anthropic and Gemini.
Streams visible text. Reasoning is retained as model-scoped native state, not exposed as visible text deltas.
Preserves tool ids, names, full arguments and error results. lite-agent still owns tool execution, permissions, hooks and concurrency.
Round-trips thinking signatures, OpenAI encrypted reasoning and assistant phase through JSON checkpoints. Native state cannot be replayed into a different provider/model.
Retains source citations as native host data without fetching their URLs. This bridge supports text, reasoning and function tools. Media/file output and provider-executed tools fail explicitly rather than being discarded or bypassing local permissions.
Aborts requests and cancels readers on cancellation/early iterator return. HTTP failures retain ProviderError.status. Retry remains an explicit lite-agent middleware choice.
Fails on vendor unsupported warnings instead of ignoring requested features. Observe other warnings through onWarning.

Provider options and storage

const provider = aiSdk(openaiModel, {
  providerOptions: { openai: { store: false } },
  onWarning: warning => console.warn(warning.type),
});

providerOptions may also be a function of the current ModelRequest. The bridge does not override a vendor's server-side storage/privacy defaults; explicitly configure OpenAI Responses store: false when appropriate. This is not redaction or an offline guarantee.

Supply context only for capabilities you actually know, such as a context window or true request-token counter. The bridge does not guess native compaction, exact counting or prompt-cache support. The direct anthropic() adapter remains available for its specialized native context management.

Validation and references

Offline tests execute installed vendor SDKs against HTTP/SSE/Bedrock binary-event fixtures, covering text/tool replay, metadata, usage, failures and no automatic retry. Providers without credentials are not claimed as live-tested. Live scripts and artifacts remain outside the repository.

#Mainstream providers through AI SDK

#Quick start

#Coverage

#Behavior