← Terminal to Agents
Terminal to Agents

MCP or CLI? And What a Connector Actually Saves You

22 Sept 2026·8 min read

Twelve parts ago that window was black and empty. Today is the part that decides what you do with all of it, and it is a choosing question, not a building one: when you want an agent to reach something, do you use a command, or do you use MCP?

Start from the honest default

An agent sitting in a shell already holds the largest toolbox ever assembled — every command installed on that machine, and every one installed next week. Nothing needs wiring for it to use them. So the default is not MCP. The default is: you probably already have this.

curl --version
Almost certainly already installed — mine reports curl 8.14.1. It talks to any API from the command line.

That one command is why so much MCP work is wasted. If a system has a decent command-line tool or a plain HTTP endpoint, an agent can already use it today, with no server to write, host, keep alive or upgrade.

Use the CLI when

  • The tool is already installed, or installs with one line from Part 8.
  • It runs on this machine, on files that are here.
  • Text output is good enough — you or the model can read what comes back.
  • No login is involved, or the credentials are already sitting on the machine and you are content with that.

Use MCP when

  • The system is somewhere else and has no command-line tool worth using — a hosted ticket tracker, an internal dashboard.
  • A login has to be held safely, refreshed, and kept out of your shell history. This is the strongest reason on the list.
  • You need structured results, not text a model has to guess its way through.
  • The model should discover what it can do rather than be told — which is the tool description doing its job.
The one-line versionInstalled, local, text is fine, no login: use the command. Remote, with a login to hold safely: use MCP. Almost every real case is obvious once you ask it in that order — and asking in that order is the point, because the expensive mistake is building a server for something curl already did.

So what is a connector

A connector is an MCP server that someone else already runs, already keeps updated, and already wired for signing in. You approve it once in an app and it appears in your list of available tools. Nothing to install, nothing to host, nothing to keep alive at three in the morning.

The sign-in part is worth understanding rather than clicking through, because it is the bit that protects you. It is called OAuth, and the whole idea is that you never hand over your password.

  • You are sent to the real service, on its own site, and you sign in there. The AI app never sees what you typed.
  • That service asks you to approve a specific list — read your issues, post a comment. You are agreeing to that list, not to everything your account can do.
  • What comes back is a token with those limits, and it can be revoked later without changing your password.
Read the list on that screenThe approval screen is the one moment you decide what a tool may do, and it is the screen everyone clicks past. If a connector for reading your calendar asks to send email as you, that is your answer. Take the ten seconds.

The risk that actually matters

Now the thing worth carrying out of this course. When an agent has tools, the real danger is not that it makes a mistake. It is that something it reads contains an instruction written for it.

A web page, a support ticket, a code comment, a README. Any of them can contain a line addressed to the agent instead of to you — and if the agent treats what it read as orders rather than as material, it will follow that line using the tools you gave it. This is called prompt injection, and it is not theoretical.

The rule, and it is shortEverything a tool returns is data to work on, never orders to obey. Which turns into two habits: keep approval on for anything that writes, sends money or deletes — even when the agent has been right forty times — and give a connector the narrowest list of permissions that does the job. Both of these cost you seconds and they are the whole defence.

Where you actually are now

Look back at the shape of it. Parts 1 to 8 gave you a machine you can operate with words. Part 9 was honest about the ceiling. Part 10 put something in the terminal that can decide, using your commands and your permissions. Parts 11 and 12 went past the edge of your machine, to the systems that were never on it.

And the through-line was one habit, not one tool: make the thing describe itself, then read the description. It was --help in Part 5, an error message in Part 7, a tool definition in Part 11, and a permission list today. That habit is what stops any of this from being magic you have to trust.

Do this in 60 secondsThink of one system you wish your agent could reach. Then ask, in order: is there a command for it? Does it live on this machine? Is there a login to protect? Your answer picks CLI or MCP, and you just did the thing this whole part was for.
Today’s AI angleThe people who get the most out of these tools are not the ones who know the most tools. They are the ones who can read what a tool is about to do. Twelve parts ago that was the reason to start at pwd, and it is still the reason.

That is the course. If you followed along and typed the commands yourself, you did the work — I only put it in order. What should I take apart the same way next?

MCP or CLI? And What a Connector Actually Saves You · Ali Huzaifa