Skip to content

Ultimate MCP Integration Doctrine

RULE: MCPs are force multipliers only when they are explicit, validated, and observable. A mysterious MCP is worse than no MCP.


1. What MCP Integration Must Achieve

An MCP integration must make these clear:

  • what tools it exposes
  • how it authenticates
  • how it is started / reached
  • what project problems it solves
  • what failure modes it introduces

2. Prefer Explicit Discovery

Always verify live tool availability rather than trusting stale documentation.

Why

MCP servers drift. Installed tools change. Old docs lie.


3. Startup & Reachability

For each MCP, document:

  • startup mode (stdio / streamable HTTP / daemon)
  • host/port if relevant
  • health-check method
  • required env vars / credentials

If you cannot prove reachability, you cannot rely on the MCP.


4. Security Requirements

MCP integrations must never silently inherit dangerous credentials or broad trust they do not need. Scope access tightly. Redact outputs where sensitive.


5. Tool Selection Discipline

Use an MCP only when it is the correct abstraction. Do not route simple local work through a remote MCP just to look sophisticated. Do not avoid an MCP when it is clearly the canonical path.


6. Failure Handling

When MCP calls fail, capture:

  • tool name
  • target MCP
  • parameters shape
  • error class
  • whether retry is safe

Do not hide MCP instability behind vague “tool failed” messages.


7. Final Rule

An MCP must reduce uncertainty, not increase it. If it adds magical hidden behavior, document and constrain it before trusting it.


Last updated: 2026-04-10
Status: ACTIVE & MANDATORY
Maintainer: sin-zeus