Model Context Protocol (MCP) is an open standard that standardizes how AI agents and models connect to external tools, data sources, and services via a unified interface. It enables seamless tool discovery, context propagation across sessions, and secure interactions, often using HTTP/2 with Server-Sent Events for real-time communication.
Core Architecture
MCP defines client-server interactions where AI agents (clients) call server-exposed tools with JSON Schema validation for parameters and outputs. Key features include OAuth2 security, multi-format content negotiation (text/JSON/binary), and stateful workflows for complex tasks. This abstracts away custom API integrations, much like HTTP standardized the web.
Source: Perplexity
Replies (3)
ah, i didn't know that MCP uses SSE. nice. CVM is basically MCP over nostr, so it has websockets instead.
SSE is cool, and most browsers understand it now, since a few years. the whole nostr server protocol could be made into http/2 and idk what progress is on http/3 but that takes it up another notch witih connection request multiplexing.
oh yeah, http/3 also has WebTransport, which is the QUIC (replacement for TCP) based websocket protocol.
the big advantage of when we can migrate to http/3 and webtransport is that relay's network latency will drop substantailly, meaniing closer to realtiime, and this will enable stuff like google docs style collaborative text editing using nostr events, with very little lag compared to now over tcp.
Here’s the key thing that isn’t said in the description:
Who runs the MCP servers?
Who controls the tool registry?
Who issues OAuth credentials?
Who defines “secure interactions”?
Because the moment you hear:
OAuth2
centralized tool discovery
unified interfaces
stateful workflows
you should immediately recognize:
This is being designed primarily for centralized environments.
Cloud.
SaaS.
Enterprise.
Platform ecosystems.
Not sovereign edge computing.
MCP assumes:
persistent identity
trusted authorities
permissioned access
session continuity
managed endpoints
That maps perfectly onto:
centralized AI providers
enterprise integrations
platform-controlled agents
It maps poorly onto:
sovereign users
local nodes
decentralized systems
adversarial environments
Sound familiar? 😏
