Migrating to v3

Version 3 of the @browser-ai packages targets Vercel AI SDK v7. The providers themselves keep the same public surface - browserAI(), transformersJS(), webLLM(), their embedding/transcription models, and the download-progress helpers are unchanged. What changes is the AI SDK version you use them with.

1. Update your dependencies

npm i ai@^7 @browser-ai/core@^3
# and/or
npm i @browser-ai/transformers-js@^3
npm i @browser-ai/web-llm@^3

AI SDK v7 requires Node.js >=22 and its packages are ESM-only. If you use @ai-sdk/react, upgrade it alongside ai.

2. Update your AI SDK call sites

These are the AI SDK v7 changes most likely to affect code that uses @browser-ai:

v6	v7
`system: "..."`	`instructions: "..."`
`stepCountIs(5)`	`isStepCount(5)`
`result.fullStream`	`result.stream`
`result.toUIMessageStream(opts)`	`toUIMessageStream({ stream: result.stream, ...opts })`
`result.toUIMessageStreamResponse()`	`createUIMessageStreamResponse({ stream: toUIMessageStream({ stream: result.stream }) })`
`result.toTextStreamResponse()`	`createTextStreamResponse({ stream: toTextStream({ stream: result.stream }) })`
`onFinish` / `onStepFinish`	`onEnd` / `onStepEnd`
`needsApproval` on `tool()`	`toolApproval: { toolName: "user-approval" }` on the call
`usage.cachedInputTokens`	`usage.inputTokenDetails.cacheReadTokens`
`usage.reasoningTokens`	`usage.outputTokenDetails.reasoningTokens`

Custom transports

Custom ChatTransport implementations are the most common place this shows up, since they call toUIMessageStream directly:

// v2 / AI SDK v6
writer.merge(result.toUIMessageStream({ sendStart: false }));

// v3 / AI SDK v7
writer.merge(
  toUIMessageStream({
    stream: result.stream,
    tools: this.tools,
    sendStart: false,
  }),
);

See the per-package useChat integration pages for full, updated transport examples.

Multi-step results

In v7, top-level result fields such as content, toolCalls, toolResults, and usage cover all steps. If you relied on final-step-only values, read them from result.finalStep instead:

const { usage, finalStep } = await generateText({ model: browserAI(), prompt });

usage.totalTokens; // all steps
finalStep.usage.totalTokens; // final step only
finalStep.finishReason;

`onChunk`

onChunk now receives every stream part, including lifecycle, boundary, finish, abort, and error parts. Always narrow on chunk.type before using a part:

onChunk: (event) => {
  if (event.chunk.type === "text-delta") {
    // ...
  }
},