API over MCP Principle

API over MCP Principle

Integration principle for AI operating systems: when an agent only needs a narrow set of actions from a service, direct API calls plus a curated reference file can be cheaper, faster, and more controllable than loading a full MCP server.

Key points

  • MCP is valuable when broad tool discovery and standardised tool calling matter, but it carries context overhead because tool definitions enter the session [src-013]
  • Direct APIs are often better for stable, high-frequency integrations where the agent needs only a small endpoint subset [src-013]
  • A curated markdown reference listing the required endpoints gives the model enough operational context without exposing the full API surface [src-013]
  • This pattern pairs naturally with the Scoped API Key Pattern: create a dedicated service account or narrowed token for the agent [src-013]
  • The principle is especially relevant across the AIOS Tier-One Domains, where each connection should be deliberate rather than maximally broad [src-013]
  • Mornati’s measurement reframes “API over MCP” as a frequency question: low-G/N services should avoid always-on schema injection, while high-G/N services can justify Native MCP [src-041]
  • Gateway MCP adds a middle option: keep structured tool outputs without loading every backend service schema upfront [src-041]

Related entities

  • ClickUp — example of using a direct API rather than the ClickUp MCP server

Related concepts

Source references

  • [src-013] Nate Herk — “Build & Sell Claude Code Operating Systems (2+ Hour Course)” (2026-05-01)
  • [src-041] Marco Mornati — “The Future of Agentic Tooling: MCP Servers vs. CLI A Data-Driven Comparison” (2026-04-27)

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 491 indexed pages and articles.

  1. Wiki concept Agentic Tooling Gives AI agents access to external systems through APIs, MCP servers, CLIs, browsers, and workflow runtimes. Related by api
  2. Wiki concept Progressive Context Loading (Skills) The three-level loading pattern Claude Code uses to keep skills lightweight. Level 1: initial search only reads the YAML frontmatter (name + Related by reference
  3. Insight AI Measurement and Experimentation How to measure AI product impact with evals, adoption metrics, online experiments, guardrails, and cost tracking Readers have engaged with this next