CLI beats MCP.

Two MCP servers taught me that agents prefer terminals: fifty years of Unix conventions, composable pipes, and --help beat custom tool schemas and protocol plumbing.

I built two MCP servers. One connects Google Search Console to my coding agent, and when it worked the first time I felt like I’d wired my house for the future.

I barely use them anymore.

What I use instead is CLIs. My agent runs gh for GitHub, wrangler for Cloudflare, ffmpeg when I need video work, a small CLI I keep around for Google Workspace. Nobody wrote an “integration” for any of this. The agent runs --help, reads the output, and gets to work.

It took me two MCP servers to figure that out.

MCP starts from the premise that an agent is a fragile API client. So you stand up a server, define tool schemas, handle an auth handshake, and keep the thing running. When it breaks, you debug a protocol. My Search Console server needs re-authing every few weeks, and the fix is me, in a terminal, running a command. The plumbing for the AI needs a human plumber.

A CLI assumes whoever’s calling it knows what they’re doing. Fifty years of Unix conventions do the rest. Flags are documented in the tool itself. Errors come back as text with an exit code. Output pipes into the next command. And an agent can compose all of it on the fly: run a command, grep the result, retry with a different flag. You don’t ship an integration; you ship a tool, and the agent writes a new integration every time it runs.

And agents grew up on terminals. Decades of shell sessions, man pages, and Stack Overflow answers are in the training data. There is no comparable corpus of people using your custom MCP server.

MCP has real uses. If your product is a SaaS with OAuth and no local anything, a hosted MCP server is a fine front door. Sandboxed environments with no shell need something like it. And schemas make sense when the model calling you is small and needs guardrails.

But if you’re a tool vendor deciding what to build for the AI wave, the order is not close. Ship a CLI. Give it a --json flag. Write the help text like you mean it. Agents can use it today, from any harness, with zero client code.

I keep my two MCP servers running, mostly as a monument. The agent walks past them to get to the terminal.