> ## Documentation Index
> Fetch the complete documentation index at: https://docs.screenpipe.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Understand AI usage and control scheduled work

> Understand hosted allowance, provider usage, background tasks, warnings, and the difference between stopping a run and disabling its schedule.

Before scheduling a report, check which AI provider will run it and how often it will analyze your history. A useful daily report can become expensive if it repeatedly rereads a large window.

## Separate the moving parts

| Part           | What it does                                            | What to check                                                                      |
| -------------- | ------------------------------------------------------- | ---------------------------------------------------------------------------------- |
| Recording      | Collects screen and audio evidence                      | Recording settings, selected devices, filters, storage, and transcription provider |
| Chat           | Runs AI when you submit a request                       | Selected model and provider                                                        |
| Scheduled task | Runs a saved instruction manually or on a schedule      | Prompt, provider, frequency, recent runs, and output                               |
| Data Sync      | Transfers configured data for a remote or team workflow | Sync settings and the [data-flow guide](/privacy-data-flow)                        |

Local capture, cloud transcription, hosted chat, and team sync are separate choices. Turning off a report's schedule does not turn off recording or cloud transcription. A local recording also does not guarantee that an AI analysis stays local: the selected provider determines where its supplied context is processed.

## Which allowance am I using?

| Selected AI route           | Where to inspect usage                                                                            |
| --------------------------- | ------------------------------------------------------------------------------------------------- |
| Screenpipe Cloud            | Screenpipe **Settings → Usage**, including available hosted limits and recent tracked sessions    |
| Your own API key            | That provider's usage and billing dashboard, plus available local run statistics                  |
| A connected AI subscription | The connected provider's account and limits                                                       |
| A local model               | Local model configuration and machine resources; verify the workflow has no separate remote calls |

Screenpipe Cloud chat and background scheduled tasks can share an account allowance. A task running in the background can therefore affect how much hosted chat remains available. External providers have their own policies and limits.

Account allowance and a model's context window measure different things. Context usage describes how much information fits into that conversation; it is not your remaining monthly balance. Local token or cost estimates may cover only tracked sessions and need not equal the provider's invoice.

Use the account's current display for its allowance and reset information. This guide does not fix a plan price, token quantity, or reset schedule that may vary by account.

## Review one task

Open **Scheduled tasks → My tasks** and select the task. Inspect its prompt, schedule, selected provider/model, and latest execution history. Then use this prompt with an assistant that can inspect that configuration:

```text theme={null}
Review [task name] without changing it. Report its provider and model,
schedule and timezone, source window per run, output destination, and recent
run status. Use available usage records to distinguish measured usage from
estimates. State any missing account or provider information.

Suggest one change that reduces repeated analysis while preserving the
output I need. Show the proposed configuration before applying it.
Do not infer remaining allowance from local token totals or assume a
subscription includes every provider call.
```

## Control runs and schedules

| Goal                              | Action                                      | Check afterward                              |
| --------------------------------- | ------------------------------------------- | -------------------------------------------- |
| Try the report once               | **Run now**                                 | Final run status and the actual saved output |
| Stop an active run                | **Stop**                                    | The execution leaves running/stopping state  |
| Prevent future automatic runs     | Turn **auto-run** off                       | The task shows **auto-run off**              |
| Stop both current and future work | Turn auto-run off, then stop the active run | Both states are confirmed                    |
| Resume the routine                | Review the schedule, then turn auto-run on  | The displayed schedule matches your intent   |

Stopping does not undo completed writes or already consumed AI usage. Auto-run off does not close a client job or pause recording. Use [scheduled task debugging](/task-troubleshooting) if an execution remains stuck.

## Warnings and spending limits

Versions with background allowance warnings can notify you when the service reports that scheduled tasks are consuming most of the hosted allowance. Delivery depends on the service response and notification preferences. These advisories are not a user-defined dollar cap or a guaranteed warning before every run.

If you require a hard spending limit, verify the controls for the provider and account you are actually using. Writing “never spend more than \$5” in an AI prompt does not create an enforced billing limit. Keep a workflow manual while its spending controls are unclear.

## Reduce repeated AI work

Start with one manual run over a short, known interval. Once accurate:

1. Match the frequency to when you need the output, such as one end-of-day report.
2. Query only the relevant date, project, and apps. Keep a checkpoint so later runs process new evidence.
3. Keep the output compact and preserve existing corrections instead of regenerating an entire knowledge base.
4. Try a less expensive available model on the same example and compare accuracy.
5. Inspect retries and failed runs. Repeated errors can consume usage without producing a useful artifact.

For example, a task that rereads the previous 24 hours every hour processes overlapping windows 24 times a day. A daily report or an incremental update reduces that repeated work; savings depend on the actual requests, model, and retries.

## When usage is unavailable

“Usage unavailable” means the app could not display the information. It does not prove that your allowance is exhausted or that usage is free.

Check the signed-in account, selected provider, and available refresh control. If the display remains unavailable, collect the app version, timestamp and timezone, error text, provider, and whether a manual chat or task works. Share that with [support](https://discord.gg/screenpipe), omitting keys and private recordings. Account entitlement and billing discrepancies require an account check; a prompt cannot repair them.

## Related guides

* [Reliable scheduled reports](/reliable-reports)
* [Build scheduled tasks](/scheduled-tasks)
* [Privacy and data flow](/privacy-data-flow)
* [Remote-support jobs](/remote-support-job-reports)