Skip to content
All guidesAI systems

MCP or CLI? Two ways an AI can use your apps

An AI that works in your apps reaches them through a command line or through MCP. Here is how each works, what each costs, and how to choose by who the AI is acting for.

6 min read

Give an AI a way into your apps and it can list open orders, update a customer record or fix a bug. The two common ways to give it that reach are a CLI and MCP. This guide explains both, compares them with real numbers and ends with a simple way to choose.

Three terms first. A CLI, or command-line interface, is a program you control by typing text commands. MCP, the Model Context Protocol, is a published standard that lets an app hand an AI a menu of tools. An API, or application programming interface, is the way software talks to an app. Both the CLI and MCP sit on top of an app's API.

How a CLI works

A CLI was built for people at a keyboard, long before AI. GitHub, where developers keep their code, has one called gh. Typing gh pr list shows the pull requests, the proposed changes waiting for review. Add a filter and only the matching lines come back.

  1. The AI writes a command in plain text, the same as a person would type.
  2. The CLI program calls the app's API.
  3. Text comes back, and the AI reads only what it asked for.

Common CLIs such as git and gh appear all over the public code that AI models learned from, so models often know the commands already. The catch is that the AI needs a computer where it can run commands, and it acts with your login and your full access.

How MCP works

Anthropic published MCP in November 2024, and AI apps including Claude and ChatGPT now support it. An MCP server is a small program that offers one app's tools to any AI app that connects to it.

  1. Connect. The server sends its menu. Each tool has a name, a description and the inputs it takes.
  2. Call. The AI picks a tool, for example list_pull_requests with the state set to open.
  3. Reply. Structured data comes back, ready for the AI to use.

Each person signs in through OAuth, a standard sign-in that gives the AI only the access that person approves. That is what makes MCP suit AI working for many people. The catch is the menu. The AI reads every tool description before it starts, and GitHub's full server offers about 90 tools, and the version in the test below offered 43. That reading is paid for in tokens, the word-pieces an AI reads and is billed for.

Side by side

  • Made for. A CLI was made for people at a keyboard. MCP was made for AI apps.
  • Learning what it can do. With a CLI, the AI reads help text when it needs it. With MCP, it gets the full tool menu up front.
  • Sign-in. A CLI uses your own login on your machine. With MCP, each user signs in with their own access.
  • What it needs. A CLI needs a computer where the AI can run commands. MCP needs an AI app that supports it, which can be a chat app in a web browser.
  • What comes back. A CLI returns text you can filter. MCP returns structured data.

The numbers from one test

In March 2026 Scalekit, a company that builds sign-in tools for AI agents, ran five read-only GitHub lookups with Anthropic's Claude Sonnet 4 model, 25 runs through gh and 25 through GitHub's MCP server. For one lookup, finding the main programming language of a project on GitHub (a project there is called a repository), the CLI used 1,365 tokens and MCP used 44,026, about 32 times more. That was the widest of the five gaps; the others ran from 4 to 20 times.

Across its 25 runs, the CLI finished all 25. MCP finished 18, and the seven misses timed out while connecting to GitHub's MCP server. Scalekit estimates 10,000 lookups a month would cost about $3.20 through the CLI and about $55.20 through MCP. That is one test by one company on one app.

Which to use

Pick by who the AI is acting for. When it works for you, on a computer that can run commands, the CLI usually uses fewer tokens. When it works for other people who each sign in, MCP keeps their access separate, so each action traces back to one person.

  • A developer asks an AI coding assistant to fix a bug on their laptop. Use the CLI. The assistant already runs commands there, and gh or git cover the job.
  • A sales team asks a chat assistant to update customer records. Use MCP. The team works in a chat app with no command line, and each person's changes should carry their own name.
  • A support assistant looks up orders for each customer who signs in. Use MCP. Each customer must see only their own orders.
  • A weekly job pulls payment totals into a spreadsheet. Use the CLI. One trusted account runs it on a schedule, and filtering the output first keeps the cost down.

Many apps offer both. GitHub has the gh CLI and an official MCP server, so you can choose per job rather than once for the whole business.

The API underneath

A CLI and an MCP server both wrap an app's API so an AI can use it. If an app has a good API, anyone can build either one. If it has no API, neither can reach it, and the AI is left clicking through screens like a person, which is slower and breaks more often.

So when you choose software for your business, ask about the API first. Both routes depend on it. Our guide to polling and webhooks shows what a good API looks like in practice, using a single order update.

Where it is heading

  • November 2025, tool search. Anthropic released a way for the AI to load only the tools a job needs. In its tests the menu shrank by 85%.
  • November 2025, code execution. The AI writes a short program that calls the tools and reads only the result. In one Anthropic example, a job fell from 150,000 tokens to 2,000.
  • July 2026, stateless MCP. Each request now stands on its own, so MCP servers are easier to run for many users at once.
  • In trial, WebMCP. A version of MCP for websites, offering their tools to an AI inside Chrome. A stable release is expected late in 2026.

Tool search and code execution cut the cost of the menu, which is MCP's main weakness today, and the other two make MCP easier to run and reach. The CLI's weaknesses, one person's login and the need for a computer that runs commands, are harder to design away.

For more on why tokens drive the bill, read our guide to what an AI holds in mind and what it costs. If you want AI connected to the systems your business already runs, our chatbots and integrations service covers that work.

Sources, checked 24 September 2026. Scalekit, MCP vs CLI benchmark (March 2026). Anthropic engineering, advanced tool use and code execution with MCP (November 2025). The Model Context Protocol blog, 28 July 2026 specification. Chrome for Developers, WebMCP.