@agent-provider/ai-sdk

An AI SDK LanguageModel provider backed by the user-controlled Agent Provider browser extension.

npm install @agent-provider/ai-sdk ai

Provider credentials remain in the extension. The page selects a configured model alias and receives structured bridge failures when permission or policy requirements are not met.

What it does

This package implements the AI SDK LanguageModelV4 interface (@ai-sdk/provider) on top of the Agent Provider extension bridge. Instead of calling a model API directly, the page sends sanitized requests to the extension over a scoped postMessage protocol; the extension applies the user's origin grants, quotas, and policy, then brokers the call to a user-configured provider using credentials the page never sees.

Page code passes only allowlisted call options across the boundary: headers and provider-controlled URLs are stripped, non-function tools are dropped, and raw chunk inclusion is disabled. The extension re-checks all of this. Failures (permission denied, quota exhausted, bridge unavailable, timeouts) surface as typed AgentProviderBridgeError instances with code, retryable, and details fields, or as AbortError DOMExceptions on cancellation.

Requirements

• A browser with the Agent Provider extension installed, and an exact-origin grant for the calling page. Without the bridge, connect() fails with a BRIDGE_UNAVAILABLE error.
• A DOM environment: the default transport uses window.postMessage, ReadableStream, crypto.randomUUID, and DOMException. For non-window contexts, supply a custom AgentProviderBridgeTransport.
• ai (AI SDK) version 7 or newer, >=7.0.34 <8.
• ESM only; the package ships a single ESM entry point with TypeScript types.

Quick start

import { generateText } from "ai";
import { createAgentProviderProvider } from "@agent-provider/ai-sdk";

const agentProvider = createAgentProviderProvider({ appName: "My App" });

const { text } = await generateText({
  model: agentProvider("default"),
  prompt: "Summarize the current page selection.",
});

The argument to the provider is a model alias configured in the extension, not a provider model ID; it defaults to "default". Streaming works through the standard AI SDK APIs (streamText and friends) with no extra configuration.

API

• createAgentProviderProvider(options?) — returns a provider callable: provider(alias?) or provider.languageModel(alias?) produces an AI SDK LanguageModelV4; provider.bridge exposes the underlying AgentProviderBridge. Options: defaultAlias, appName, clientId, connectTimeoutMs (default 3000), requestTimeoutMs (default 90000), a pre-built bridge, or a custom transport.
• createAgentProviderModel(options?) — convenience escape hatch returning a LanguageModel directly; accepts alias plus the bridge options above.
• AgentProviderBridge — lower-level session object. connect() opens the session and resolves the extension-reported BridgeCapabilities; requestPermission() and refreshPermission() drive the consent flow; requestToolApproval() and reportToolExecution() support the tool-use contract; snapshot holds the latest capabilities; dispose() tears down all pending work.
• AgentProviderBridgeError — structured failure with code, retryable, and details.
• AgentProviderBridgeTransport / WindowAgentProviderTransport — transport interface and the default window-based implementation.
• Re-exported protocol types: BridgeCapabilities, BridgeLimits, PermissionState.

Links

• Repository — extension, threat model, and implementation status.
• @agent-provider/protocol — the wire protocol and capability types this bridge speaks.

License

CC0-1.0 OR Unlicense.