OpenAI added event-triggered tasks to ChatGPT for Gmail, Slack and GitHub. The model now launches a workflow the moment a matching email lands, a Slack post appears, or a pull-request event occurs. This eliminates hourly polling and delivers instant, context-aware automations for the tools we use every day.

Why the shift matters

Previously, ChatGPT’s automation tools checked for new data only once an hour. Users often waited for a draft email, a Slack summary, or a code review. By listening directly to platform events, OpenAI turned ChatGPT into a reactive assistant that drafts replies, summarizes discussions, or creates changelogs the instant a trigger fires.

How the feature works

The new tasks appear in the ChatGPT web interface and mobile app; they do not run in the desktop client, the Codex command-line tool, or any IDE extensions. A task consists of several fields:

  • Source – the Gmail account, Slack channel or GitHub repository that emits the event.
  • Event – the action to listen for, such as a new email, a Slack message in a chosen channel, or a pull-request review, comment or merge.
  • Filter – optional criteria that narrow the trigger, like sender address, email subject, Slack author, thread ID or GitHub label.
  • Output – the result format, for example a concise summary or a draft ready for human approval.
  • Forbidden action – an explicit instruction telling the model not to perform high-risk operations such as merging code, pushing commits or sending messages.
  • Dedup key – a unique identifier that prevents duplicate runs.

A single task can listen to several triggers, but it cannot mix an event trigger with a timed schedule. When multiple matching events arrive close together, the system may coalesce them into one run; a single event does not always guarantee a separate execution.

Setting up the triggers

  • Gmail – Choose “new message” as the event and add filters on sender or subject to keep the automation focused.
  • Slack – Select “new message” in a specific channel; filters can target the author or a thread. Reactions and direct messages are not supported as triggers.
  • GitHub – Pick pull-request activities such as reviews, comments or merges. The associated GitHub app must have repository access for the task to fire.

Safety first

OpenAI warns against letting a freshly created task modify data straight away. Start with read-only actions—have the model generate a draft email, a summary comment, or a proposal document—so you can review before anything is sent or merged. Listing forbidden actions adds an extra guardrail.

Testing is essential before relying on a task in production:

  1. Positive test – fire a matching event and verify that the task starts.
  2. Negative test – send an unrelated event and confirm that nothing runs.
  3. Coalescing test – trigger two events in quick succession to see if they merge appropriately.
  4. Permission test – revoke the task’s access and ensure it fails safely rather than proceeding with incomplete rights.

OpenAI suggests limiting the number of sources or event types in early runs, adding complexity only after the automation proves reliable.

Who gains, who watches

Takeaway

OpenAI’s event-triggered tasks turn ChatGPT from a periodic assistant into a real-time collaborator on Gmail, Slack and GitHub. Reacting instantly to incoming data speeds up turnarounds and reduces manual stitching of information, but it also demands careful filtering, explicit safety constraints and thorough testing.