@effect-uai/ai-sdk

Vercel AI SDK compatibility for effect-uai. Keep your @ai-sdk/react useChat frontend exactly as it is and serve it from an effect-uai agent loop instead of streamText.

Two functions bridge the two worlds:

• Messages.decodeMessages turns the UIMessage[] a useChat client POSTs into effect-uai HistoryItems.
• UIMessageStream.toUIMessageStream projects the loop's InteractionEvent stream onto the AI SDK UI Message Stream protocol (v1), emitting SSE.Events.

The package owns no HTTP layer: it produces a Stream<SSE.Event> and the required responseHeaders, so it drops into any server.

Install

pnpm add @effect-uai/ai-sdk

A chat route

decodeMessages in, run your loop, toUIMessageStream out. Provide your provider layer so the stream has no remaining requirements, then hand the bytes to whatever server you run.

import * as Messages from "@effect-uai/ai-sdk/Messages"
import * as UIMessageStream from "@effect-uai/ai-sdk/UIMessageStream"
import * as SSE from "@effect-uai/core/SSE"
import { Stream } from "effect"

export async function POST(request: Request): Promise<Response> {
  const { messages } = await request.json()
  const history = Messages.decodeMessages(messages)

  const events = agent(history).pipe(Stream.provide(OpenAILayer)) // your loop, unchanged

  const body = events.pipe(
    UIMessageStream.toUIMessageStream(crypto.randomUUID()),
    SSE.toBytes,
    Stream.toReadableStream,
  )
  return new Response(body, { headers: UIMessageStream.responseHeaders })
}

With @effect/platform

Same stream, wrapped in an HttpServerResponse instead of a web Response:

import { HttpServerResponse } from "@effect/platform"

const body = events.pipe(UIMessageStream.toUIMessageStream(id), SSE.toBytes)
return HttpServerResponse.stream(body, { headers: UIMessageStream.responseHeaders })

Coverage

Outbound, text and reasoning deltas, tool calls (input streaming plus resolved output), and refusals map to their protocol parts. Alongside the loop's InteractionEvents, toUIMessageStream also accepts dataPart and messageMetadata emissions, so callers can interleave custom typed data (data-<name>, optionally transient) and settled per-message metadata:

events.pipe(
  Stream.map((e) => (Metrics.isThroughput(e) ? dataPart("metrics", e, { transient: true }) : e)),
  UIMessageStream.toUIMessageStream(id),
)

Inbound, decodeMessages reconstructs text, image (file with an image/* media type), and tool parts: an assistant tool call becomes a function_call item plus a function_call_output once the client carries its result. Non-image files and unresolved tool states are dropped.