A Grain of Salt

MCP vs. CLI + Skill: When to Use Which

· Teddy Aryono

Overview

A key architectural decision in agentic workflows is whether to use an MCP server (e.g., GitHub MCP) or a locally installed CLI paired with a Skill (e.g., gh CLI + GitHub skill).

MCP (e.g., GitHub MCP)

CLI + Skill (e.g., gh CLI + GitHub skill)

Decision Matrix

ScenarioPrefer
Cloud/sandboxed agent (no shell)MCP
Simple CRUD on well-defined resourcesMCP
Complex multi-step shell workflowsCLI + Skill
Need to pipe/compose with other toolsCLI + Skill
Auth managed externallyMCP
Agent runs locally with full env accessEither (CLI often simpler)
Portability across environmentsMCP
Debugging / observabilityCLI (easier to see raw output)

The Real Differentiator

Execution environment + composability.

For Claude Code / Local Agent Workflows

When enterprise customers run Claude Code locally, CLI + Skill is often the right default because:

  1. The agent already has shell access
  2. The developer likely has the CLI authenticated
  3. The Skill gives you a repeatable way to encode team conventions — branching strategy, PR templates, etc. — that a generic MCP server won’t know about

#ai #ai-agents #engineering

Reply to this post by email ↪