As described informally in #179

I propose a new MCP protocol element:

Event

As MCP becomes the senses of our LLMs, we're lacking the ability to receive feedback from our environment. We need to be able to SMELL!

Events should be as simple as possible to keep our LLMs understanding them completely.

public interface Event {
   topic: string // the topic name of the event
   ts: number // timestamp of when the event occurred for real
   body: object // the body of the event
}

export interface Topic {
  name: string
  description: string
  bodySchema: {
    type: "object";
    properties?: { [key: string]: object };
    required?: string[];
  }
 }
}

export interface ListTopicsRequest {}

export interface ListTopicsResponse {
  topics: Topics[]
}

export interface SubscribeRequest {
  topic: string
}
export interface SubscribeResponse {}
export interface UnsubscribeRequest {
  topic: string
}

export interface ListEventsRequest {
  linger?: number // the amount of time it waits for events to collect before returning
}
export interface ListEventsResponse {
  events: Event[] // the events generated since the last call to ListEvents
}

We could also support an event JSON rpc message if the connection support streaming (STDIO or HTTP streaming) otherwise its long polling for the agents. That's ok. We've done it forever with CometD.

With the inclusion of this functionality, LLMs could be told IF THIS THEN THAT type logic that uses tools and subscriptions and gets feedback through events.

I want to call out the obvious ethical concerns of giving an AI LLM the ability to react to environment and tools to affect said environment before it's brought up.

I would love to know how best we can make sure those concerns are addressed.