ChatGPT MCP Events: Webhook Triggers Arrive on a Draft Spec

What changed

ChatGPT no longer has to wait for a user to type something before it does work inside a connected app. It can now subscribe to events emitted by a plugin's Model Context Protocol (MCP) server and take action when one of those events arrives 1. In practice, an assistant could watch an external system and respond when something changes in it, rather than acting only when prompted.

The catch is what this is built on. OpenAI itself describes the underlying specification as a draft 1. Delivery works through webhooks only 12. The feature arrived alongside OpenAI's DevDay 2026 announcements, where a computer-use launch in the Agents API also came with constraints that builders will need to weigh 2.

A partial implementation of an unfinished standard

The specification comes from the MCP Triggers and Events Working Group, which was chartered in March 2026 2. Its proposal for events, submitted through the protocol's enhancement process (an SEP), is still listed at the "ideating" stage 2. That is an early point in the process. It suggests the design questions are still open, not that the work is nearly finished.

ChatGPT implements only part of that draft 2. The supported path is webhook delivery: the MCP server pushes an event to an endpoint, and ChatGPT reacts 12. Several mechanisms are absent [2]:

• polling, where the client checks for new events on its own schedule
• streaming delivery
• handling for gap notifications

The last omission deserves the most attention. Gap notifications typically exist to tell a consumer that it may have missed events, for example after an outage or a dropped connection. Without them, a webhook-only system depends on every delivery succeeding. If one fails, nothing in this setup necessarily tells the automation that it is now working from an incomplete picture. For low-stakes workflows that may be acceptable. For anything where a missed event has consequences, builders would likely need to add their own reconciliation or retry logic.

The computer-use fine print

The second issue concerns a different product. The computer-use capability in OpenAI's Agents API ships with data residency limited to the United States and offers no Zero Data Retention option 2. OpenAI is presenting managed infrastructure as a selling point: let the company run the agent environment so you don't have to 2.

For many enterprise buyers, and especially for teams with workloads in the EU, residency and retention controls are often prerequisites for procurement rather than optional extras. A US-only footprint with no zero-retention mode could put the offering out of reach for those teams until the options broaden 2. That limits how far the managed-infrastructure pitch reaches for exactly the customers most likely to want it.

How the two reports differ

Both reports agree on the core facts: ChatGPT can act on MCP events, delivery is webhook-only, and the spec is not final 12. They differ mainly in emphasis. One frames the news around a capability shift, with ChatGPT moving from reactive chat toward watching apps and acting, while noting that OpenAI itself labels the foundation a draft 1. The other reads more like a builder's checklist. It spells out the working group's status, lists the delivery modes that are missing, and connects the events story to the computer-use compliance limits as part of a single decision: build now or wait 2.

Reading the move

This looks like a familiar platform strategy. OpenAI is shipping early against an emerging standard to establish a position, and openly accepting that the standard may change underneath it. Shipping a narrow, webhook-only slice is a reasonable way to do that. It is the simplest delivery model to support, and it lets developers start experimenting with event-driven agents now.

The risk falls on builders. A spec at the "ideating" stage can change in ways that break integrations 2. A working group chartered only months ago has not had time to settle hard problems such as missed-event recovery. Teams that build production automations on this surface today should treat them as provisional. That means adding their own safeguards for missed events and planning for spec revisions.

The practical split is fairly clear. Prototypes, internal tools, and workflows where an occasional dropped trigger is tolerable are good candidates for experimenting now. Regulated, EU-bound, or mission-critical automations have reasons to wait, both for the events specification to mature and for OpenAI to expand residency and retention options on computer use 2. For now, the main thing ChatGPT's new ability to watch and act offers developers is an early look at where agent platforms are heading.