MCP Events Draft Spec: What ChatGPT Supports After DevDay 2026

ChatGPT starts listening to MCP servers

OpenAI's DevDay 2026 delivered a wide set of agent and developer updates. For anyone building MCP servers, one change stands out: ChatGPT can now subscribe to events pushed from connected applications. The company said ChatGPT supports the proposed MCP Events specification on all plans. As a result, plugins can start automations when something happens in an external app, such as a new message, a comment or a status change 15. The event took place on September 29, 2026, and support requires MCP 2.0 at protocol version 2026-07-28 2.

The practical gain is that agents no longer have to poll. A server can notify ChatGPT when something occurs, and ChatGPT then acts according to instructions the user has already set 12. This is a meaningful step from reactive chat toward background automation.

A spec that isn't finished yet

The main caveat is that MCP Events is not a finished standard. The MCP Triggers and Events Working Group was chartered in March 2026, and its first specification enhancement proposal for events is still listed at the ideation stage 1. ChatGPT implements only part of the draft.

According to implementation guidance, ChatGPT accepts webhook delivery only 12. Polling and streaming delivery are not supported, and neither are the draft's gap and terminated notifications 1. Gap and terminated notifications are the mechanisms that would tell a client it missed events or that a subscription ended.

Server builders must do the following [2]:

• Implement events/list, events/subscribe and events/unsubscribe on the same authenticated endpoint as their other MCP methods.
• Declare an events capability in server/discover.
• Verify callback endpoints.
• Sign every delivery using Standard Webhooks HMAC.

Servers also need persistent storage for subscriptions and outbound HTTPS access to callback URLs 1.

This amounts to a vendor shipping ahead of a standards body. That pattern can speed adoption, but it carries risk. If the working group revises the draft, developers who build to ChatGPT's subset may need to rework their servers. The missing gap notifications deserve particular attention: without them, a server has no protocol-level way to signal dropped events. Builders may need their own reconciliation logic for anything where reliability matters.

The wider plugin and agent push

MCP Events is one part of a larger plugin expansion. Developers can now build sidebar experiences, interactive conversation panels and custom file viewers inside ChatGPT 5. On the API side, the Agents API gained computer use, which lets applications operate software through graphical interfaces 5. It also absorbs Codex's multi-agent capabilities, tool search, tool calling and context compaction, with OpenAI running the execution infrastructure. These features are available via the API and in Codex and ChatGPT Work on selected plans 5. Other announcements included GPT-6.1 Sol and a Decisions API 5.

OpenAI managing execution reduces the operational burden for developers. It also concentrates more of the agent stack on OpenAI's side, which teams will need to weigh against their own governance requirements. The announcements covered here do not detail where that managed execution runs or what data-residency options exist. Teams with compliance obligations should confirm those details directly before relying on the service.

Codex moves to the cloud

Codex received four announcements [3]:

• Cloud environments: Codex can run tasks in a prepared environment with repositories, dependencies and tools, and work continues while a laptop is closed 34.
• Refreshed CLI: The Codex CLI adds voice input and an /agents interface for delegating and monitoring multiple tasks 5.
• Code review: A new workflow analyzes diffs and flags potential issues.
• Codex Security Cloud: A scanner that checks repositories and new commits, investigates findings, removes duplicates and prepares fixes 5.

Details vary across reports. On plan availability, one account lists Plus, Pro, Business, Healthcare, Education and Enterprise, and says Free and Go are excluded 3. Another names Pro, Business, Healthcare, Education and Enterprise without mentioning Plus 4.

Code review is described in different ways:

• Generally available for GitHub 3
• Working with both GitHub pull requests and GitLab merge requests 5
• An experience inside the ChatGPT desktop app, with summaries and diff exploration 4

Codex Security Cloud is characterized as a research preview 3. The older Codex Cloud (Legacy) still handles code review and the Linear and GitHub integrations 3.

The takeaway

The common thread is unattended work. Codex runs in the cloud, Security Cloud scans repositories without supervision, and MCP Events lets ChatGPT react to outside triggers. OpenAI is building agents that act without a human prompting every step.

For MCP developers, the advice is to build event support now, but conservatively. Target ChatGPT's webhook-only subset, sign and verify everything, and assume the specification will change.