You already know: task-shaped tools, user-scoped permissions, timeouts, sandboxing and untrusted outputs (Tools & protocols); processes, pipes and standard input/output (D1 Operating systems); HTTP, streaming responses and APIs (D1 Networking).
Why this matters for Atlas. Without a standard, every AI application integrates every tool its own way: N apps × M tools integrations, each with its own bugs. The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in 2024 and now supported across major AI apps, IDEs and agent frameworks, that defines how an application discovers and calls tools, reads data and fetches prompt templates from a server. Write the Atlas course-search server once; every MCP-capable client can use it. This lesson builds a working MCP-style server and client over a real pipe, explains the primitives and transports, and spends real time on the part that bites in production: security.
1. Hosts, clients and servers
The question. Who talks to whom in MCP?
Start from nothing. Three roles: