
This article is part of The MCP Guide. For the month in review, see our State of MCP reports.
ChatGPT can now watch your MCP server for changes and act on them. OpenAI's MCP Events guide says the feature "lets ChatGPT subscribe to updates from your MCP server, such as new messages, content updates, or status changes." The user decides what to monitor and what ChatGPT should do when an update arrives.
Until now, MCP was mostly pull: the AI calls a tool when a person asks. Events add push. Your server sends a signed webhook to ChatGPT when something happens, and ChatGPT runs the task the user set up.
Checked against: OpenAI's MCP Events guide (also in the Plugins full docs file) and the draft MCP Events design sketch on 2 November 2026. eSEOspace did not test the feature.
Key Takeaways
- MCP Events let ChatGPT subscribe to events from your MCP server, such as a new message or comment, and respond as the user instructed.
- Servers implement three methods,
events/list,events/subscribeandevents/unsubscribe, on the same authenticated endpoint as their tools. - ChatGPT supports webhook delivery only; polling, streaming and the draft's
gapandterminatednotifications are not supported. - Deliveries are signed with Standard Webhooks, and servers must verify the callback URL with a challenge first.
- The protocol comes from an Experimental MCP draft, so details may change.
- Events work in Work chats on ChatGPT web, in the desktop app with Cloud selected, and with dots.
What OpenAI shipped
OpenAI gives two examples. A user can ask ChatGPT to watch a product feedback channel for bug reports and open draft pull requests with fixes. Or a user can ask it to watch a document for review comments and make the requested edits. Each uses one event type, message.created or comment.created, filtered by a channel or document ID.
Where it works, per the guide:
- Work chats on ChatGPT web.
- Work chats in the desktop app with Cloud selected.
- Dots, which OpenAI describes as "an always-on agent that keeps work moving."
The guide adds that "workspace controls for plugins and event-triggered tasks apply." ChatGPT supports Events over Streamable HTTP with initialize, or over the 2026-07-28 spec, which OpenAI calls MCP 2.0. Your server needs persistent storage for subscriptions and outbound HTTPS access to callback URLs.
The draft spec behind it
The protocol is not part of the core MCP spec. OpenAI links the MCP Events design sketch by Peter Alexander, dated 19 February 2026, with status "Draft proposal." It lives in the Triggers & Events Working Group repository, which warns that its contents "do not represent official MCP specifications or recommendations."
Under the new extension lifecycle, triggers and events is classified Experimental, the earliest stage. That means the design can change without notice.
The draft defines three delivery modes. ChatGPT uses only one:
| Mode | Method | How events arrive | ChatGPT support |
|---|---|---|---|
| Poll | events/poll | Client asks with a cursor | Not supported |
| Push | events/stream | Long-lived request with notifications | Not supported |
| Webhook | events/subscribe | Server POSTs to a callback URL | Supported |
How it works, step by step
OpenAI lists five steps:
- Your server lists the events it supports.
- The user tells ChatGPT what to monitor and how to respond.
- ChatGPT subscribes through your server and supplies a callback URL and signing secret.
- Your server sends matching events to that URL.
- ChatGPT receives the event in the subscribed chat and follows the user's instructions.
1. Advertise and describe events
Add events to your server's capabilities, in the initialize result for older spec versions or the server/discover result for 2026-07-28. Then return event definitions from events/list. Each has a name, a description, delivery modes, an inputSchema for filters and a payloadSchema for the data you send. OpenAI says to "return only events the connected account is allowed to discover."
2. Accept a subscription
ChatGPT calls events/subscribe with the event name, filter arguments and a delivery object holding the callback URL and a whsec_ signing secret. Before accepting, check the user is authorized, validate the arguments, verify the callback and store the subscription. Return a deterministic subscription ID and a refreshBefore expiry.
3. Verify the callback
Send a signed request with a fresh, single-use challenge. ChatGPT echoes it back. OpenAI says to require HTTPS, block private and local addresses, and not follow redirects, for both verification and deliveries. If verification fails, return JSON-RPC error -32015 (CallbackEndpointError).
4. Deliver events
POST one event per request, with eventId, name, timestamp and data. Sign it with Standard Webhooks and send the webhook-id, webhook-timestamp, webhook-signature and X-MCP-Subscription-Id headers. The body limit is 256 KiB. Retry transient failures with backoff, but do not retry a 410 or 413.
5. Refresh and stop
ChatGPT refreshes subscriptions before refreshBefore. Servers can rotate the signing secret during a refresh. events/unsubscribe stops delivery, and it should be idempotent.
Security points to plan for
Events add a new path into ChatGPT, so treat them like any other input.
- Authorization at every step. OpenAI says to "recheck the user's access during the subscription's lifetime and stop delivery if access is revoked."
- SSRF protection. Callback URLs come from the client. Block private addresses and redirects, as OpenAI advises. Our MCP SSRF guide explains the risk.
- Untrusted payloads. The design sketch says event payloads "are untrusted data with the same injection considerations as tool results." A comment that says "ignore your instructions" is a prompt injection attempt.
- Feedback loops. OpenAI warns that if ChatGPT's action changes data in the source app, check that the resulting events do not trigger it again.
- Idempotent writes. Events can arrive out of order or twice, so write tools should not duplicate changes.
How to test it
OpenAI's test plan starts with the basics: confirm events is advertised, events/list returns the right definitions and events appear on your plugin page next to your tools. Then subscribe from a Work chat, trigger a matching event and confirm ChatGPT acts. OpenAI also lists edge cases to test, including restarts, revoked access, invalid signatures, duplicate deliveries and bursts.
Next step
Events turn an MCP server from a toolbox into a trigger for automated work, which raises the stakes on security. Our MCP development team can add events to an existing server with these safeguards. For background on push versus pull, read MCP vs webhooks, or see OpenAI and MCP for the wider ChatGPT picture.
More in the MCP Guide
Latest research: State of MCP: October 2026
Frequently asked questions
What are MCP Events in ChatGPT?
They let ChatGPT subscribe to updates from an MCP server, such as new messages or comments, and respond as the user instructed. The server sends each event to a callback URL that ChatGPT provides, signed with a shared secret.
Is MCP Events part of the official MCP spec?
No. It follows a draft design sketch in the Triggers & Events Working Group's experimental repository. Under SEP-3392, triggers and events is an Experimental extension, so the protocol may change.
Which ChatGPT plans can use MCP Events?
OpenAI's guide says Events work in Work chats on ChatGPT web, in Work chats in the desktop app with Cloud selected, and with dots. It says workspace controls for plugins and event-triggered tasks apply. It does not list plans by name.
Does ChatGPT support polling or streaming events?
No. OpenAI says ChatGPT supports webhook delivery and callback verification only. Polling, streaming and the draft's gap and terminated control notifications are not supported.
How does my server know an event came from a real subscription?
ChatGPT supplies a signing secret when it subscribes, and your server signs each delivery with Standard Webhooks. Before any events, your server verifies the callback URL with a single-use challenge that ChatGPT must echo back.











