MCP is becoming event-driven
MCP Events is now in beta. Agents can subscribe to things like new emails, GitHub issues or incidents and react when they happen, instead of waiting for a prompt.

MCP Events is now in beta. Agents can subscribe to things like new emails, GitHub issues or incidents and react when they happen, instead of waiting for a prompt.

Something interesting happened in the MCP ecosystem this week.
MCP Events is now in beta.
The spec behind it, SEP-3415, is final, and the extension is officially in beta. I posted about it on X and a lot of you had questions, so here's the longer version.
Until now, almost every MCP interaction has followed a request-response pattern.
An agent needs something, it calls a tool, the server responds, and the agent continues.
That works well when someone is asking. But think about how most real-world applications work.
Things happen without us asking for them.
An email arrives. A payment succeeds. Someone opens a GitHub issue. A production server goes down.
Wouldn't it be better if an agent could subscribe to these events and get notified when they happen?
That's exactly what MCP Events makes possible.
It does. Things like notifications/resources/updated. But they were never built for this:
Events fixes all three.
The flow is simple:
events/list). Each event type says what arguments it takes and what data it sends.Not every setup can do the same thing. A serverless server can't hold open connections. A client behind a firewall can't receive a callback. So Events gives you three options, and the server picks the ones that fit:
| Mode | How it works | Good for |
|---|---|---|
| Poll | The client checks for new events at an interval the server suggests | Serverless servers, simple clients |
| Push | The client keeps a connection open and events stream in | Local servers, clients that stay connected |
| Webhook | The server sends signed events to a URL the client registers | Remote servers, clients that can't stay connected |
This is the part I'm most excited about.
Imagine an agent that watches your inbox and prepares replies when important emails arrive.
An agent that watches your GitHub repositories and starts looking into new issues.
Or an agent that monitors production and starts investigating the moment something goes wrong.
None of these need someone sitting there typing a prompt. They run on their own and wake up when something happens.
I think that's a big step for MCP. We've spent a lot of time making it easy for agents to do things in other apps. But being useful isn't only about asking and acting. It's also about knowing when something happens.
Receiving an event does not mean the agent is allowed to act on it.
The spec is clear about this:
MCP Events is an extension, so both the client and the server have to opt in. It needs MCP protocol version 2026-07-28 or later. ChatGPT already documents support for it, and you can check the client matrix to see where others are.
If you want to go deeper:
It's still in beta, and I'm really looking forward to seeing how clients and servers start adopting it.
At MewCP we sit between agents and the apps they use, so this is something we're watching closely. More on that soon.