Daily

Get a daily list of failed predictions

Each morning you see which runs failed in the past day, which model they used and the error, so you can fix them before rerunning.

Runs: Daily
Apps: Replicate
AI: Not needed
Access: API key

Send this to your assistant

Create a ClickRuntime automation: every morning at 7am, list my Replicate predictions from the last 24 hours that failed and email me the model, start time and error for each.

Paste it into your assistant. It tells you what it will send or change before it runs live.

Add to ChatGPT Add to Claude

First time? Adding ClickRuntime to your assistant takes about two minutes.

Why use it

Why teams use this automation

Teams that run image, audio or language models on Replicate pay for compute by the second, so a prediction that fails after running for minutes still costs money. When a pipeline calls the API thousands of times a day, failures hide in the volume: a bad input format, a model update that changed behavior or a timeout that keeps repeating.

Each day, ClickRuntime lists the predictions from the last 24 hours, keeps the ones that failed and emails you a report. It shows which models failed, the error messages and how long each prediction ran, so you can see where money was wasted and where a pattern needs fixing.

It is most useful to developers and product teams with an application that sends many requests to Replicate and cannot review each one.

What you get out of it

Wasted compute is visible
A failed prediction that ran for ninety seconds before erroring has still been billed. Seeing run times next to the failures shows how much money disappeared into broken requests. You can see the money leaking.
Repeated errors point to a bug
The same error message appearing fifty times is almost certainly a bug in your input handling, not bad luck. Grouping failures makes that obvious while the cause is recent. The fix is usually small once found.
Model changes are noticed
When a model owner publishes a new version that rejects previously valid input, your failure count jumps overnight. The daily report shows the jump on the first day instead of after a week.
Spend drivers get named
Listing which models account for most of the failed runs focuses cost-cutting on the right place. You can switch model, tighten validation or add retries only where they matter. Savings come from fixing the largest causes first.
Your users stop being the monitoring system
A customer whose image never generated might not report it. A daily list of failures lets you find and fix the problem for everyone, including those who never complained. Fewer people hit the same problem silently.

How it works

What happens, step by step

Every morning, ClickRuntime goes through all your Replicate predictions from the last 24 hours.
It keeps only the ones that failed.
If any failed, you get one email listing each model, the time it ran and the error, so you can fix it before rerunning.

Before it runs live

Connect Replicate

What you get

RunsDaily

Get a daily list of failed predictionstoday, 9:00Succeeded

Email · Failed predictionsnightmareai/real-esrgan at 2026-10-18T22:14: CUDA out of memory

Build your own

Need it to work differently?

This template is a starting point. Copy the prompt and change it, or describe a different automation from scratch, with your own conditions, more apps or an AI step. Your assistant writes it, tests it and runs it on ClickRuntime.

Your own rules
Conditions, thresholds, filters and exceptions, written the way you'd explain them to a colleague.
Any number of apps
One automation can read from one app and update several others. Over 248 apps are ready, and anything with an API works too.
AI where it helps
Summaries, classification, extraction and drafts use ClickRuntime's built-in AI, with no extra key.
On a schedule or on each event
Runs as often as every 5 minutes, or the moment an app sends an event, and keeps track of what it already handled.

Changes you can ask for

Run it at a different time, more or less often, or only on weekdays
Add an AI step, for example to summarize, sort or draft a reply
Add another app, for example also send it to Slack
Combine it with other Replicate steps into one automation

Under the hood

The code that runs this automation

This is the Python script that runs each time the automation starts. Your AI assistant writes it for you when you send the prompt, and you allow each step in your assistant. To change how the automation works, describe the change to your assistant and it updates the script for you. For every function it can use, see the script reference.

failed_prediction_daily_report.pyexample

from datetime import datetime, timedelta, timezone
from clickruntime import on_trigger, credentials, notify, http

@on_trigger
def run(event):
    headers = {"Authorization": f"Bearer {credentials.secret('REPLICATE_API_TOKEN')}"}
    since = (datetime.now(timezone.utc) - timedelta(hours=24)).isoformat()
    url, failed = "https://api.replicate.com/v1/predictions", []
    while url:
        page = http.get(url, headers=headers, params={"created_after": since}).json()
        failed += [p for p in page["results"] if p["status"] == "failed"]
        url = page.get("next")
    lines = [f"{p['model']} at {p['created_at'][:16]}: {p.get('error')}" for p in failed]
    if lines:
        notify.email_me(f"{len(lines)} failed Replicate predictions", "\n".join(lines))
    return {"failed": len(lines)}

FAQ

Common questions

What does Replicate count as a failed prediction?

A prediction with status failed, meaning the model raised an error or the run could not complete. Canceled predictions have their own status and are not counted as failures. The error text from Replicate is included in the report.

Does the report show how much the failures cost?

It shows prediction run times, which determine billing for most models. Exact prices depend on the hardware each model uses, so costs are approximate. For exact numbers, compare with your usage page in Replicate. Use it to compare days against each other.

Can it report only for a specific model?

Yes. Predictions record the model or version used, so the report can be limited to those you care about. This is helpful when you run many models but only one is business critical, and the others fail often for testing reasons.

What permissions does it need?

API key. Nothing else in your accounts is used.

Does it need AI?

No. It runs without AI, so it doesn't use your plan's AI allowance.

How many runs does it use?

One run each time it runs (daily). Checks that find nothing to do and stop right away aren't counted.

Can I change how it works?

Yes. Describe the change to your AI assistant, for example a different schedule, channel or filter. It updates the automation, tests it and tells you what changed.

Daily