Are we overengineering AI agents, or am I just lazy?

MCP makes sense when you have multiple agents sharing the same remote service. Real use case, no debate.

But on your laptop? Running a server so your agent can query a database is insane. Just use psql.

We overcomplicated this.

It's faster. It's simpler. Claude already knows how to use these tools.

I kept a few MCP servers where no CLI alternative exists yet. Context7 for example. But that's the exception, not the rule.

Then I just wrote one line in my CLAUDE.md: "Use CLI tools."

psql instead of a Postgres MCP server.
gh instead of a GitHub MCP server.
curl instead of a search MCP server.
agent-browser instead of Playwright MCP server.

Done. No server running. No config. No protocol.