An agent built with the AI SDK deploys like any other application, with one command. Time spent waiting on a model is part of your plan, and an agent that goes past its plan is throttled, never billed extra.
The example agent’s code is in swarza/examples.
waitUntil and Next.js after() keep running for up to 15 minutes after the last request: saving a summary, calling a webhook, finishing a tool call. If work reaches the limit, the application’s logs say so.
// app/api/report/route.ts
import { after } from "next/server";
export async function POST(req: Request) {
const chat = await req.json();
after(() => summarize(chat));
return Response.json({ queued: true });
}Follow the AI SDK’s resume guide: the page reconnects with useChat({ resume: true }) and gets the rest of the answer. The chunks wait in the application’s memory, so there is no Redis to run.
// components/chat.tsx useChat({ id, resume: true }); // app/api/chat/[id]/stream/route.ts const stream = await context.resumeExistingStream(id); if (!stream) return new Response(null, { status: 204 }); return new Response(stream, { headers });
Create a database with one command and save each chat when its answer ends, so a reload, or a new deployment, shows the whole conversation.
$ swarza databases create notes --app my-app
import { connect } from "@tursodatabase/serverless";
const db = connect({
url: process.env.DATABASE_URL,
authToken: process.env.DATABASE_AUTH_TOKEN,
});
const notes = await db.all("SELECT * FROM notes");An MCP server over Streamable HTTP is an ordinary application. A scheduled job can run an agent every morning, for up to 5 or 15 minutes depending on the plan.
// swarza.json { "scheduledJobs": [ { "schedule": "0 7 * * 1-5", "function": "jobs/digest.mjs" } ] } // jobs/digest.mjs export default async function () { const { text } = await generateText({ model, prompt }); await sendDigest(text); }
// app/api/chat/route.ts
export async function POST(req: Request) {
const { messages } = await req.json();
const result = streamText({ model, messages, tools });
return createUIMessageStreamResponse({
keepAliveMs: 15_000,
stream: toUIMessageStream({ stream: result.stream }),
});
}Where each one runs an agent, how you pay for it, and what happens to a long answer.
| Swarza | VercelFunctions | CloudflareAgents | |
|---|---|---|---|
| Entry price | €5 / month (Starter) | $20 / month per developer seat (Pro) | $5 / month minimum (Workers Paid) |
| How you pay | One flat monthly price | Active CPU, memory time and invocations | Requests and CPU time. Agents also by wall-clock time |
| Past the included usage | Throttled, never billed extra | Billed, uncapped. Pausing at a budget is opt-in | Billed. Budget alerts don’t pause or cap usage |
| Longest streamed answer | No limit, with a keep-alive every 15 s | 800 s (1,800 s in beta) | No limit while connected |
| Work after the response | Up to 15 min after the last request | Within the same limit | waitUntil: 30 s. Alarms: 15 min |
| Durable workflows | Not yet, planned | Workflow SDK, billed per event | Cloudflare Workflows |
| Resumable streams | The AI SDK’s resume, without Redis | You add Redis, or use Workflows | Built in |
| Conversation storage | Turso databases, included | Partners (Neon, Upstash, Supabase) | SQLite in each agent |
| Languages | JavaScript and TypeScript (Node.js, Bun) | JavaScript, TypeScript, Python | JavaScript and TypeScript |
| WebSockets | No | Beta, closed at the function’s limit | Yes, the main transport |
| EU region | Frankfurt, always | Frankfurt and others, if chosen. Default: Washington, D.C. | Agents’ data, by jurisdiction. Workers: Enterprise add-on |
Other providers’ prices are in US dollars, excluding tax, as their pages show them.
