Scheduled tasks

Beta feature: tasks that run on their own: on a schedule, via webhook or by email. Setup, triggers, results, notifications and limits.

With scheduled tasks, an assistant (or a model) works on its own without you opening a chat. A task starts on a schedule, when another system calls a webhook URL, or when an email arrives at the task's address. You find them in the sidebar under Scheduled.

Note

Scheduled tasks are currently a beta feature. The Scheduled entry in the sidebar therefore carries a Beta label. Support enables the feature per workspace. If the Scheduled entry is missing from your sidebar, the feature is not active for your workspace yet. Contact support in that case.

Create a task

On the Scheduled page, click Create task. Alternatively, open the Scheduled tasks card in the assistant editor and click Add task. By default, all members have the Create automations permission. Admins can restrict it, see Permissions.

FieldMeaning
NameUp to 120 characters.
What should set it off?Schedule, Webhook or Email. You cannot change the trigger after creating the task. Delete the task and create a new one to switch.
InstructionsWhat the task should do on every run, up to 20,000 characters. The placeholder {{now}} is replaced with the current time.
Assistant or modelEither an assistant whose prompt, tools and knowledge the task uses, or a model with the tools you switch on in the selection menu. For an assistant you need edit rights. If your workspace only allows assistants, you have to pick one.
Time zone, Frequency, TimeSchedule trigger only.
Allowed sendersEmail trigger only, see below.

A new task is active right away. It runs as your account: it uses your model access and your own connections, but only the connectors that apply to the task (see Connectors in tasks).

Triggers

Schedule

Choose one of the frequencies Daily, Every Monday to Friday, Weekly on (one weekday) or Monthly on (one day of the month), plus a time and the time zone. With Custom… you set:

  • Weekly on several weekdays, for example Monday and Thursday.
  • Monthly on a day from 1 to 28, on the last day of the month, or on a weekday such as the last Friday.

In the interface you set one time per task, so the schedule starts a task at most once a day. Run now is exempt from that, and you can use it at any time on top of the schedule. The card shows the next run, for example "Next run: today at 08:00". After an outage, the task catches up on a missed slot once, not on every missed slot. A task never overlaps itself: if the previous run is still going, the next one waits.

Webhook

The task gets its own webhook URL with a secret token. Another system starts it with a POST request and a JSON body. The response 202 with a runId confirms that the run is queued. The content of the request reaches the model as a separate message and counts as received data, not as an instruction. Use the Idempotency-Key header to detect duplicate calls: a second call with the same key returns the same runId.

The URL is the secret: anyone who knows it can start the task. With Renew address in the edit dialog you generate a new URL. The old one stops working, so enter the new one everywhere the old one was configured.

Email

The task gets its own email address on the domain @agents.meingpt.com. Every mail to that address starts a run, including BCC and forwarded mail. Sender, subject and body reach the model as a separate message and count as received data, not as an instruction. MeinGPT truncates the body at 64 KB and marks it with [truncated]. Up to 10 attachments with a combined 20 MB are passed along. MeinGPT tells the model about files that are too large or cannot be retrieved instead of failing the run. With Renew address you generate a new address, and the old one stops working.

You decide who may send mail in the Allowed senders field:

  • Everyone in the workspace (default): only senders whose email address belongs to a member of your workspace start the task.
  • Specific addresses or domains: only the entries you type into the input field, separated by commas, start the task. An entry is a full address or a domain with a leading @, for example anna@company.com, @partner.com. @partner.com allows every address on exactly that domain, but not subdomains such as @mail.partner.com.
  • Any sender: every mail to the task's address starts a run, whoever sent it. Only workspace admins can choose this option, after a security prompt. The change is recorded in the audit log.

The check uses the sender address (From) of the incoming mail. SPF and DKIM are not checked.

Attention

Any sender is a deliberate risk. Anyone who knows the task's address can start a run. The mail body reaches the model as data, but it can still influence the model, and the task works with the tools and access of the person who runs it. Only the secret address, the task's hourly limit and the per-sender cap protect it then. Choose this option only for tasks where a stranger's run cannot do harm, and renew the address if it becomes known.

Attention

Automatically forwarded mail: when a mail-server rule forwards messages from third parties, the original author usually stays in the From field. If that sender is not a workspace member and not on the list, the mail is dropped silently and starts no run, even though your mail server reports successful delivery. In that case enter the original senders or their domains under Specific addresses or domains. If the mail comes from many different senders, say a forwarded shared mailbox, Any sender is the only option.

If a forwarded mail contains a block such as "Forwarded message" or "Weitergeleitete Nachricht", the task shows the model the original author as the sender. The sender check still applies to the actual sender of the mail.

Runs and results

Every run creates its own hidden chat. Open the task to see the history. A run has one of these states:

  • Queued: the run is waiting.
  • Running …: the run is working.
  • Succeeded: the run finished and reported success.
  • Failed: the run was aborted or reported an error.
  • Nothing to do: the model reported that no action was needed.
  • Unverified: the model delivered no closing report, so the result is not confirmed. This does not count as a failure.

Clicking a run opens its chat. For failed runs, and for other editors, the Run result dialog opens instead, with the summary or error and the files the run created. Use Load older runs to see earlier runs.

MeinGPT deletes the body of received emails from the run data after 30 days. Sender and subject remain.

Notifications

  • Bell: every finished run creates an entry with the task, its state and a summary.
  • Email notification: when you edit a task, you decide when an email with a summary of the result goes to up to 20 recipients: Never (default), After every run, Only with a result (success only, not "Nothing to do") or Only on failure. The email contains the short summary the assistant writes at the end of the run (at most 2,000 characters) as plain text without formatting, not the full result text from the chat. Use the button to open the task and the link to the full transcript in the email to reach the formatted result in MeinGPT. Files up to a combined 15 MB are attached to the mail. Larger files come as a link that is valid for 7 days.

Manage tasks

The card and the detail page offer these actions:

  • Run now starts a test run, even for a paused or switched-off task. Use it to check a fix before you switch the task back on. For email and webhook tasks, the model knows that nothing was received.
  • Pause and Activate switch the task off and on.
  • Edit opens the dialog. Maximum runs per hour and the email notification are set here, not when creating a task.
  • Delete removes the task. Chats it already created are kept. If you delete an assistant, its tasks are deleted too.

The Scheduled page offers a search by name or assistant and a status filter (All, Active, Paused, Off). Paused tasks always sort first.

Connectors in tasks

A task does not inherit the connectors from your regular chats. Which connectors a run can use depends on how the task is set up:

  • Task with an assistant: Only the connectors enabled in the assistant apply. The selection under Connectors in the task dialog is locked in this case. If, for example, Microsoft Outlook is missing from the assistant, the run cannot use Outlook even though Outlook works in your chat.
  • Task with a model: The connectors you switch on in the task dialog via the "+" menu in the Instructions field, under Connectors, apply.

In both cases the run uses your own connections to the services. If a run cannot reach a service, first check that the connector is selected in the assistant or in the task dialog, then start a test run with Run now.

Limits

  • A run ends after 9 minutes and uses at most 20 tool steps. Failed runs are not retried automatically.
  • Only one run of a task runs at a time.
  • By default, a task allows 60 runs per hour, counted across all triggers including Run now. For email and webhook tasks you change the value in the edit dialog (1 to 1000).
  • A single email sender can start the same task at most 25 times per hour.

Credits and automatic pausing

Scheduled tasks use the workspace's credits, see Credits. MeinGPT pauses a task automatically when

  • 5 runs in a row fail, or
  • the workspace's credits are used up.

You recognize paused tasks by the yellow dot next to Scheduled in the sidebar, the notice on the page and the Paused status. After even one failed run, the detail page shows a warning. You usually also receive an email. If the creator is no longer active, the emails go to the workspace admins. Fix the cause and switch the task back on with Activate. If credits ran out, an admin has to top up first.

When nothing happens

Mails and calls that don't start a task often report no error back. Check these points:

  • An email starts no run: Is the sender allowed? For forwarded mail this is usually the cause, see above. A domain entry covers exactly that domain, not its subdomains. Is the task active and not paused? Has the hourly limit or the limit of 25 mails per sender been reached? Both lead to silent dropping. The mail does not bounce, and the task history shows no entry.
  • A webhook answers with 404: The URL is unknown, the task is paused or switched off, or the feature is not active for the workspace. 429 means the hourly limit is reached.
  • A run shows "Unverified": The model did not confirm a result. Open the run's chat and check what happened.
Was this page helpful?