Replicate webhooks that finish what they start.

The timing problem

Replicate completed-run callbacks via AnyHook

Pass the AnyHook inbound URL as the `webhook` parameter on your prediction. Your handler downstream gets all the time it needs to fan out.

import Replicate from "replicate";
import { verifyWebhook } from "@anyhook/verify";

const replicate = new Replicate();

// 1. Start the prediction with AnyHook as the callback target.
await replicate.predictions.create({
  version: "...",
  input: { prompt: "..." },
  webhook: `https://in.anyhook.net/${userSlug}/${appSlug}`,
  webhook_events_filter: ["completed"],
});

// 2. Handler — runs after AnyHook has already acked Replicate.
export async function POST(req: Request) {
  if (!(await verifyWebhook(req, process.env.ANYHOOK_SIGNING_SECRET!))) {
    return new Response("invalid", { status: 401 });
  }

  const { id, output } = await req.json();
  await persistOutput(id, output);   //  3s
  await notifyUser(id);              //  2s
  await chainNextModel(output);      // 45s — fine, we have 5 min
  return new Response("ok", { status: 200 });
}

Questions teams ask

Does this work with all Replicate model types?

Yes — image, video, audio, language. The webhook contract is identical. AnyHook treats the payload as opaque JSON and forwards it intact.

Replicate already has retries. Why do I need AnyHook?

Replicate retries on transport failures, but its window is short and it won't retry your downstream business logic — only the webhook delivery. AnyHook gives you 10 retries with exponential backoff against your handler, plus one-click replay weeks later.

Can I see the prediction output without checking Replicate again?

Yes. The full webhook payload — including the `output` URLs — is stored encrypted in your AnyHook event log and inspectable in the dashboard.