Back to AI/ML Overview
🔌 Knowledge Site · Model Context Protocol · Identity, Authorization, Governance

MCP EngineeringA protocol standardizes the call. It does not grant the permission.

The lets many agent applications discover and call one shared capability. This site starts from one plain function that reads a roof assessment and climbs, one step at a time, to a remote used by four teams: when a function, an HTTP API or is the right choice, how identity and authorization really work, who may publish and approve a tool, and how a platform team keeps a shared capability safe, observable and governable.

🎯The problem

An insurer's assessment service is wanted by intake, underwriting, claims and analytics agents. Giving every agent database credentials is unsafe; wrapping the service in does not make it safe either, because the protocol carries the call, not the permission. engineering decides when a protocol server is worth it, and then puts identity, scoped authorization, audit and ownership around it.

🔑The question every page asks
Function, HTTP API or : which is the simplest thing that is safe here, and who is accountable for it? One synthetic assessment capability, from a Python function to a governed remote shared by four teams.

🗺️What is inside

From function to protocol

A plain function, , client and server, then : host, client, server, tools, resources and prompts, per the current revision.

The decisions

Local or remote, the database decision (same-process code, a Java service or an adapter), and when sharing across teams justifies a protocol.

Identity and security

User versus agent identity, OAuth scopes, enterprise-managed authorization, no token passthrough, and the threat model: injection, tool poisoning, confused deputy.

Running it for many teams

Governance and ownership, the publishing lifecycle, audit, resilience, testing and an enterprise reference architecture.

🖱️How to read it

  • New to : read in order. Each chapter starts in plain English with one tiny example.
  • Architects: jump to the database decision, authorization or governance. The rigorous part sits in folded blocks.
  • Nine interactive labs, among them five architectures for one capability, a who-can-call-which-tool matrix and a prompt-injection red team.
  • The example is synthetic and its agent is a scripted stand-in. Results show how the mechanism works, not how good any model is.