MCP Tools vs Skills with Scripts - Why Simpler Might Be Better

I've had a patchy experience with MCP tools. Some work brilliantly. Most don't, or at least not reliably enough that I reach for them by default.

The MCP problem

MCP servers get built to solve a specific problem for someone. They work well in that context. Then they get shared, and people try using them in ways the author didn't anticipate. Things break. There are exceptions — some MCP tools are robust and well-designed — but recently I find myself treating MCP as a last resort rather than a first choice.

Maybe it's just immaturity. MCP as a protocol is still finding its feet.

What actually works

CLI tools, on the other hand, are incredibly mature. The gh CLI, glab, git itself — these are battle-tested, well-documented, predictable. They've been used in anger by thousands of people across wildly different contexts. The rough edges got smoothed out years ago.

When I build a skill around a CLI tool, I know it's going to work. The AI agent calls a stable, well-understood command. The output is consistent. Error handling is documented. I'm building on solid ground.

Is MCP actually needed?

I keep wondering if skills with scripts — or skills wrapping mature CLI tools — might be a more stable platform for AI agents than MCP. Simpler architecture, fewer moving parts, less abstraction between the agent and the actual work being done.

Not saying MCP has no place. But right now, for most tasks, a good CLI tool wins.