vercelAI

vercelAI is an HTTP route adapter. Point it at a route handler that already uses the Vercel AI SDK; AgentsKit sends the conversation to that route and consumes its streamed response.

import { vercelAI } from '@agentskit/adapters'

const adapter = vercelAI({
  api: '/api/chat',
})

This adapter does not wrap a LanguageModel object. Your Vercel AI SDK model and provider configuration stay inside the route handler.

#Route handler

The adapter posts messages, systemPrompt, and any declared AgentsKit tools. For a conversation without tool results, a minimal Next.js route can keep the existing model boundary and return the AI SDK UI Message Stream:

import { streamText } from 'ai'

type TextMessage =
  | { role: 'user'; content: string }
  | { role: 'assistant'; content: string }
  | { role: 'system'; content: string }

type AgentsKitRouteBody = {
  messages: TextMessage[]
  systemPrompt?: string
}

export async function POST(request: Request) {
  const { messages, systemPrompt }: AgentsKitRouteBody = await request.json()
  const result = streamText({
    model: 'anthropic/claude-sonnet-4.5',
    system: systemPrompt,
    messages,
  })

  return result.toUIMessageStreamResponse()
}

toUIMessageStreamResponse() sets x-vercel-ai-ui-message-stream: v1; the adapter consumes its SSE text and reasoning deltas until [DONE]. A route that returns plain text is also supported, but it does not carry UI-stream metadata.

Keep tool execution inside the route if you need it: configure AI SDK tools there, run them under your server policy, and return their final streamed text. The current adapter does not preserve AgentsKit toolCallId values or assistant tool-call parts across this HTTP boundary, so it cannot round-trip an AgentsKit tool history into AI SDK structured parts. vercelAI therefore declares no tool capability by default. Use a direct provider adapter when the AgentsKit runtime must own the tool loop.

#Options

Use same-origin routes when possible. For a remote route, authenticate it through headers, enforce authorization server-side, and avoid exposing reusable secrets in browser bundles.

#Why use this adapter

• Keep the Vercel AI SDK route and provider choices you already operate.
• Add an AgentsKit runtime or another AgentsKit surface without rewriting the model boundary.
• Change the provider behind the route without changing AgentsKit consumers.
• Migrate incrementally instead of coupling application code to a second provider SDK.

#Related

• Providers overview · Migrating from Vercel AI SDK
• AI SDK UI stream protocol